# ================================================================= # # Merged and Integrated Script for Audio/MIDI Processing and Rendering (Stereo Enhanced) # # This script combines two functionalities: # 1. Transcribing audio to MIDI using two methods: # a) A general-purpose model (basic-pitch by Spotify). # b) A model specialized for solo piano (ByteDance). # - Includes stereo processing by splitting channels, transcribing independently, and merging MIDI. # 2. Applying advanced transformations and re-rendering MIDI files using: # a) Standard SoundFonts via FluidSynth (produces stereo audio). # b) A custom 8-bit style synthesizer for a chiptune sound (updated for stereo output). # # The user can upload a Audio (e.g., WAV, MP3), or MIDI file. # - If an audio file is uploaded, it is first transcribed to MIDI using the selected method. # - The resulting MIDI (or an uploaded MIDI) can then be processed # with various effects and rendered into audio. # #================================================================ # Original sources: # https://huggingface.co/spaces/asigalov61/ByteDance-Solo-Piano-Audio-to-MIDI-Transcription # https://huggingface.co/spaces/asigalov61/Advanced-MIDI-Renderer #================================================================ # Packages: # # sudo apt install fluidsynth # # ================================================================= # Requirements: # # pip install gradio torch pytz numpy scipy matplotlib networkx scikit-learn # pip install piano_transcription_inference huggingface_hub # pip install basic-pitch pretty_midi librosa soundfile # # ================================================================= # Core modules: # # git clone --depth 1 https://github.com/asigalov61/tegridy-tools # # ================================================================= import io import os import hashlib import time as reqtime import copy import shutil import librosa import pyloudnorm as pyln import soundfile as sf import torch import ffmpeg import gradio as gr from dataclasses import dataclass, fields # ADDED for the parameter object # --- Imports for Vocal Separation --- import torchaudio from demucs.apply import apply_model from demucs.pretrained import get_model from demucs.audio import convert_audio from src.piano_transcription.utils import initialize_app from piano_transcription_inference import PianoTranscription, utilities, sample_rate as transcription_sample_rate # --- Import core transcription and MIDI processing libraries --- from src import TMIDIX, TPLOTS from src import MIDI from src.midi_to_colab_audio import midi_to_colab_audio # --- Imports for General Purpose Transcription (basic-pitch) --- import basic_pitch from basic_pitch.inference import predict from basic_pitch import ICASSP_2022_MODEL_PATH # --- Imports for 8-bit Synthesizer & MIDI Merging --- import pretty_midi import numpy as np from scipy import signal # ================================================================================================= # === Hugging Face SoundFont Downloader === # ================================================================================================= from huggingface_hub import hf_hub_download import glob # --- Define a constant for the 8-bit synthesizer option --- SYNTH_8_BIT_LABEL = "None (8-bit Synthesizer)" # ================================================================================================= # === NEW: Central Parameter Object === # ================================================================================================= @dataclass class AppParameters: """A dataclass to hold all configurable parameters for the application.""" # This provides type safety and autocomplete, preventing typos from string keys. # Input files (not part of the settings panel) input_file: str = None batch_input_files: list = None # Global Settings s8bit_preset_selector: str = "Custom" separate_vocals: bool = False # --- Advanced Separation and Merging Controls --- enable_advanced_separation: bool = False # Controls visibility of advanced options separate_drums: bool = True separate_bass: bool = True separate_other: bool = True transcribe_vocals: bool = False transcribe_drums: bool = False transcribe_bass: bool = False transcribe_other_or_accompaniment: bool = True # Default to transcribe 'other' as it's most common merge_vocals_to_render: bool = False merge_drums_to_render: bool = False merge_bass_to_render: bool = False merge_other_or_accompaniment: bool = False enable_stereo_processing: bool = False transcription_method: str = "General Purpose" # Basic Pitch Settings onset_threshold: float = 0.5 frame_threshold: float = 0.3 minimum_note_length: int = 128 minimum_frequency: float = 60.0 maximum_frequency: float = 4000.0 infer_onsets: bool = True melodia_trick: bool = True multiple_pitch_bends: bool = False # Render Settings render_type: str = "Render as-is" soundfont_bank: str = "None (8-bit Synthesizer)" render_sample_rate: str = "44100" render_with_sustains: bool = True merge_misaligned_notes: int = -1 custom_render_patch: int = -1 render_align: str = "Do not align" render_transpose_value: int = 0 render_transpose_to_C4: bool = False render_output_as_solo_piano: bool = False render_remove_drums: bool = False # 8-bit Synthesizer Settings s8bit_waveform_type: str = 'Square' s8bit_pulse_width: float = 0.5 s8bit_envelope_type: str = 'Plucky (AD Envelope)' s8bit_decay_time_s: float = 0.1 s8bit_vibrato_rate: float = 5.0 s8bit_vibrato_depth: float = 0.0 s8bit_bass_boost_level: float = 0.0 s8bit_smooth_notes_level: float = 0.0 s8bit_continuous_vibrato_level: float = 0.0 s8bit_noise_level: float = 0.0 s8bit_distortion_level: float = 0.0 s8bit_fm_modulation_depth: float = 0.0 s8bit_fm_modulation_rate: float = 0.0 s8bit_adaptive_decay: bool = False s8bit_echo_sustain: bool = False s8bit_echo_rate_hz: float = 5.0 s8bit_echo_decay_factor: float = 0.6 s8bit_echo_trigger_threshold: float = 2.5 # --- Anti-Aliasing & Quality Parameters --- s8bit_enable_anti_aliasing: bool = True # Main toggle for all new quality features s8bit_use_additive_synthesis: bool = False # High-quality but CPU-intensive waveform generation s8bit_edge_smoothing_ms: float = 0.5 # Mild smoothing for standard waveforms (0 to disable) s8bit_noise_lowpass_hz: float = 9000.0 # Lowpass filter frequency for noise s8bit_harmonic_lowpass_factor: float = 12.0 # Multiplier for frequency-dependent lowpass filter s8bit_final_gain: float = 0.8 # Final gain/limiter level to prevent clipping # ================================================================================================= # === Helper Functions === # ================================================================================================= def one_pole_lowpass(x, cutoff_hz, fs): """Simple one-pole lowpass filter (causal), stable and cheap.""" if cutoff_hz <= 0 or cutoff_hz >= fs/2: return x dt = 1.0 / fs rc = 1.0 / (2 * np.pi * cutoff_hz) alpha = dt / (rc + dt) y = np.empty_like(x) y[0] = alpha * x[0] for n in range(1, len(x)): y[n] = y[n-1] + alpha * (x[n] - y[n-1]) return y def smooth_square_or_saw(note_waveform, fs, smooth_ms=0.6): """Short triangular smoothing to soften sharp edges (simple anti-alias-ish).""" if smooth_ms <= 0: return note_waveform kernel_len = max(1, int(fs * (smooth_ms/1000.0))) # triangular kernel k = np.convolve(np.ones(kernel_len), np.ones(kernel_len)) # triangle shape length=2*kernel_len-1 k = k / k.sum() # pad and convolve y = np.convolve(note_waveform, k, mode='same') return y def additive_bandlimited_waveform(wave_type, freq, t, fs, max_harmonics_cap=200): """ Simple additive band-limited generator: - saw: sum_{n=1..N} sin(2π n f t)/n - square: sum odd harmonics sin(2π n f t)/n N chosen so n*f < fs/2. This is heavier but yields much less aliasing. """ nyq = fs / 2.0 max_n = int(nyq // freq) if max_n < 1: return np.zeros_like(t) max_n = min(max_n, max_harmonics_cap) y = np.zeros_like(t) if wave_type == 'Sawtooth': # saw via Fourier series for n in range(1, max_n + 1): y += np.sin(2*np.pi * n * freq * t) / n # normalization to [-1,1] y = - (2/np.pi) * y else: # square n = 1 while n <= max_n: y += np.sin(2*np.pi * n * freq * t) / n n += 2 y = (4/np.pi) * y # clip tiny numerical overshoot y = np.clip(y, -1.0, 1.0) return y def safe_tanh_distortion(x, strength): """Milder soft clipping: scale then tanh, with adjustable drive.""" # make strength between 0..1 typical; map to drive factor drive = 1.0 + strength * 4.0 return np.tanh(x * drive) / np.tanh(drive) def prepare_soundfonts(): """ Ensures a default set of SoundFonts are downloaded, then scans the 'src/sf2' directory recursively for all .sf2 files. Returns a dictionary mapping a user-friendly name to its full file path, with default soundfonts listed first in their specified order. Downloads soundfont files from the specified Hugging Face Space repository to a local 'src/sf2' directory if they don't already exist. Returns a list of local paths to the soundfont files. """ SF2_REPO_ID = "asigalov61/Advanced-MIDI-Renderer" SF2_DIR = "src/sf2" # This list is now just for ensuring default files exist # {"Super GM": 0, "Orpheus GM": 1, "Live HQ GM": 2, "Nice Strings + Orchestra": 3, "Real Choir": 4, "Super Game Boy": 5, "Proto Square": 6} DEFAULT_SF2_FILENAMES = [ "SGM-v2.01-YamahaGrand-Guit-Bass-v2.7.sf2", "Orpheus_18.06.2020.sf2", "Live HQ Natural SoundFont GM.sf2", "Nice-Strings-PlusOrchestra-v1.6.sf2", "KBH-Real-Choir-V2.5.sf2", "SuperGameBoy.sf2", "ProtoSquare.sf2" ] # Create the target directory if it doesn't exist os.makedirs(SF2_DIR, exist_ok=True) # --- Step 1: Ensure default SoundFonts are available --- print("Checking for SoundFont files...") for filename in DEFAULT_SF2_FILENAMES: local_path = os.path.join(SF2_DIR, filename) # Check if the file already exists locally to avoid re-downloading if not os.path.exists(local_path): print(f"Downloading '{filename}' from Hugging Face Hub...") try: # Use hf_hub_download to get the file # It will be downloaded to the specified local directory hf_hub_download( repo_id=SF2_REPO_ID, repo_type='space', # Specify that the repository is a Space filename=f"{filename}", # The path to the file within the repository local_dir=SF2_DIR, # local_dir_use_symlinks=False # Copy file to the dir for a clean folder structure ) print(f"'{filename}' downloaded successfully.") except Exception as e: print(f"Error downloading {filename}: {e}") # If download fails, we might not be able to use this soundfont # --- Step 2: Scan the entire directory for all .sf2 files --- print(f"Scanning '{SF2_DIR}' for all .sf2 files...") all_sfs_map = {} # Use glob with recursive=True to find all .sf2 files in subdirectories search_pattern = os.path.join(SF2_DIR, '**', '*.sf2') for full_path in glob.glob(search_pattern, recursive=True): # Create a user-friendly display name, including subfolder if it exists relative_path = os.path.relpath(full_path, SF2_DIR) display_name = os.path.splitext(relative_path)[0].replace("\\", "/") # Use forward slashes for consistency all_sfs_map[display_name] = full_path # --- Step 3: Create the final ordered dictionary based on priority --- ordered_soundfont_map = {} # Create display names for default files (filename without extension) default_display_names = [os.path.splitext(f)[0] for f in DEFAULT_SF2_FILENAMES] # Separate other files from the default ones other_display_names = [name for name in all_sfs_map.keys() if name not in default_display_names] other_display_names.sort() # Sort the rest alphabetically # Add default soundfonts first, maintaining the order from DEFAULT_SF2_FILENAMES for name in default_display_names: if name in all_sfs_map: # Check if the file was actually found by the scanner ordered_soundfont_map[name] = all_sfs_map[name] # Add all other soundfonts after the default ones for name in other_display_names: ordered_soundfont_map[name] = all_sfs_map[name] return ordered_soundfont_map # ================================================================================================= # === 8-bit Style Synthesizer (Stereo Enabled) === # ================================================================================================= def synthesize_8bit_style(*, midi_data: pretty_midi.PrettyMIDI, fs: int, params: AppParameters, progress: gr.Progress = None): """ Synthesizes an 8-bit style audio waveform from a PrettyMIDI object. This function generates waveforms manually instead of using a synthesizer like FluidSynth. Includes an optional sub-octave bass booster with adjustable level. Instruments are panned based on their order in the MIDI file. Instrument 1 -> Left, Instrument 2 -> Right. Now supports graded levels for smoothing and vibrato continuity. This enhanced version includes advanced anti-aliasing and quality features to produce a cleaner, less harsh sound. """ total_duration = midi_data.get_end_time() # Initialize a stereo waveform buffer (2 channels: Left, Right) waveform = np.zeros((2, int(total_duration * fs) + fs)) num_instruments = len(midi_data.instruments) # Phase tracking: main oscillator phase for each instrument osc_phase = {} # Vibrato phase tracking vibrato_phase = 0.0 # Retrieve anti-aliasing settings, using getattr for backward compatibility use_aa = getattr(params, 's8bit_enable_anti_aliasing', False) # --- Move progress tracking to the note level --- # 1. First, collect all notes from all instruments into a single list. all_notes_with_instrument_info = [] for i, instrument in enumerate(midi_data.instruments): # --- Panning Logic --- # Default to center-panned mono pan_l, pan_r = 0.707, 0.707 if num_instruments == 2: if i == 0: # First instrument panned left pan_l, pan_r = 1.0, 0.0 elif i == 1: # Second instrument panned right pan_l, pan_r = 0.0, 1.0 elif num_instruments > 2: if i == 0: # Left pan_l, pan_r = 1.0, 0.0 elif i == 1: # Right pan_l, pan_r = 0.0, 1.0 # Other instruments remain centered # Store each note along with its parent instrument index and panning info for note in instrument.notes: all_notes_with_instrument_info.append({'note': note, 'instrument_index': i, 'pan_l': pan_l, 'pan_r': pan_r}) # Initialize oscillator phase for each instrument osc_phase[i] = 0.0 # Independent phase tracking for each instrument # 2. Create an iterable for the main note-processing loop. notes_iterable = all_notes_with_instrument_info total_notes = len(notes_iterable) # 3. Wrap this new iterable with tqdm if a progress object is available. if progress and hasattr(progress, 'tqdm'): notes_iterable = progress.tqdm( notes_iterable, desc="Synthesizing Notes...", total=total_notes ) # 4. The main loop now iterates over individual notes, not instruments. for item in notes_iterable: note = item['note'] i = item['instrument_index'] pan_l = item['pan_l'] pan_r = item['pan_r'] freq = pretty_midi.note_number_to_hz(note.pitch) note_duration = note.end - note.start num_samples = int(note_duration * fs) if num_samples <= 0: continue t = np.arange(num_samples) / fs # --- Graded Continuous Vibrato --- # This now interpolates between a fully reset vibrato and a fully continuous one. # Use accumulated phase to avoid vibrato reset per note vib_phase_inc = 2 * np.pi * params.s8bit_vibrato_rate / fs per_note_vib_phase = 2 * np.pi * params.s8bit_vibrato_rate * t continuous_vib_phase = vibrato_phase + np.arange(num_samples) * vib_phase_inc # Weighted average of the two phase types final_vib_phase = ( per_note_vib_phase * (1 - params.s8bit_continuous_vibrato_level) + continuous_vib_phase * params.s8bit_continuous_vibrato_level ) vibrato_lfo = params.s8bit_vibrato_depth * np.sin(final_vib_phase) # Update the global vibrato phase for the next note if num_samples > 0: vibrato_phase = (continuous_vib_phase[-1] + vib_phase_inc) % (2 * np.pi) # --- Waveform Generation with FM --- fm_lfo = params.s8bit_fm_modulation_depth * np.sin(2 * np.pi * params.s8bit_fm_modulation_rate * t) modulated_freq = freq * (1 + fm_lfo) # --- Waveform Generation (with Anti-Aliasing options) --- use_additive = use_aa and getattr(params, 's8bit_use_additive_synthesis', False) if use_additive and params.s8bit_waveform_type in ['Square', 'Sawtooth']: note_waveform = additive_bandlimited_waveform(params.s8bit_waveform_type, freq, t, fs) else: # --- Waveform Generation (Main Oscillator with phase continuity) --- phase_inc = 2 * np.pi * (modulated_freq + vibrato_lfo) / fs phase = osc_phase[i] + np.cumsum(phase_inc) if num_samples > 0: osc_phase[i] = phase[-1] % (2 * np.pi) # Store last phase if params.s8bit_waveform_type == 'Square': note_waveform = signal.square(phase, duty=params.s8bit_pulse_width) elif params.s8bit_waveform_type == 'Sawtooth': note_waveform = signal.sawtooth(phase) else: # Triangle (less prone to aliasing) note_waveform = signal.sawtooth(phase, width=0.5) if use_aa and params.s8bit_waveform_type in ['Square', 'Sawtooth']: edge_smooth_ms = getattr(params, 's8bit_edge_smoothing_ms', 0.5) note_waveform = smooth_square_or_saw(note_waveform, fs, smooth_ms=edge_smooth_ms) # --- Bass Boost (Sub-Octave Oscillator) --- if params.s8bit_bass_boost_level > 0: bass_freq = freq / 2.0 # Only add bass if the frequency is reasonably audible if bass_freq > 20: # Bass uses a simple square wave, no vibrato, for stability bass_phase_inc = 2 * np.pi * bass_freq / fs bass_phase = np.cumsum(np.full(num_samples, bass_phase_inc)) bass_sub_waveform = signal.square(bass_phase, duty=0.5) # Mix the main and bass waveforms. # As bass level increases, slightly decrease main waveform volume to prevent clipping. main_level = 1.0 - (0.5 * params.s8bit_bass_boost_level) note_waveform = (note_waveform * main_level) + (bass_sub_waveform * params.s8bit_bass_boost_level) # --- Noise & Distortion (Reordered and Improved) --- if params.s8bit_noise_level > 0: raw_noise = np.random.uniform(-1, 1, num_samples) * params.s8bit_noise_level if use_aa: noise_cutoff = getattr(params, 's8bit_noise_lowpass_hz', 9000.0) raw_noise = one_pole_lowpass(raw_noise, cutoff_hz=noise_cutoff, fs=fs) note_waveform += raw_noise # --- Distortion (Wave Shaping) --- if params.s8bit_distortion_level > 0: if use_aa: note_waveform = safe_tanh_distortion(note_waveform, params.s8bit_distortion_level) else: # Original harsher distortion # Using a tanh function for a smoother, "warmer" distortion note_waveform = np.tanh(note_waveform * (1 + params.s8bit_distortion_level * 5)) # --- ADSR Envelope Generation (with improvements) --- start_amp = note.velocity / 127.0 envelope = np.zeros(num_samples) min_attack_s = 0.001 # 1 ms minimum attack to prevent clicks if params.s8bit_envelope_type == 'Plucky (AD Envelope)': attack_samples = max(int(min_attack_s * fs), min(int(0.005 * fs), num_samples)) # --- Adaptive Decay Logic --- # This ensures short staccato notes have the same initial decay rate # as long notes, fixing the perceived low volume issue. if params.s8bit_adaptive_decay: # 1. Calculate the "ideal" number of decay samples based on the user's setting. ideal_decay_samples = int(params.s8bit_decay_time_s * fs) if ideal_decay_samples <= 0: ideal_decay_samples = 1 # Avoid division by zero. # 2. Create the full, "ideal" decay curve from peak to zero. ideal_decay_curve = np.linspace(start_amp, 0, ideal_decay_samples) # 3. Determine how many decay samples can actually fit in this note's duration. actual_decay_samples = num_samples - attack_samples if actual_decay_samples > 0: # 4. Take the initial part of the ideal curve, sized to fit the note. num_samples_to_take = min(len(ideal_decay_curve), actual_decay_samples) # Apply the attack portion. envelope[:attack_samples] = np.linspace(0, start_amp, attack_samples) # Apply the truncated decay curve. envelope[attack_samples : attack_samples + num_samples_to_take] = ideal_decay_curve[:num_samples_to_take] # --- Original Decay Logic (Fallback) --- else: decay_samples = min(int(params.s8bit_decay_time_s * fs), num_samples - attack_samples) envelope[:attack_samples] = np.linspace(0, start_amp, attack_samples) if decay_samples > 0: envelope[attack_samples:attack_samples+decay_samples] = np.linspace(start_amp, 0, decay_samples) else: # Sustained envelope = np.linspace(start_amp, 0, num_samples) if use_aa and num_samples > 20: # Add a tiny release fade to prevent clicks release_samples = int(min(0.005*fs, num_samples // 10)) if release_samples > 0: envelope[-release_samples:] *= np.linspace(1.0, 0.0, release_samples) # --- Hybrid Note Smoothing (Proportional with an Absolute Cap) --- # This improved logic calculates the fade duration as a percentage of the note's # length but caps it at a fixed maximum duration. This provides the best of both worlds: # it preserves volume on short notes while ensuring long notes have a crisp attack. if params.s8bit_smooth_notes_level > 0 and num_samples > 10: # 1. Define the maximum allowable fade time in seconds (e.g., 30ms). # This prevents fades from becoming too long on sustained notes. max_fade_duration_s = 0.03 # 2. Calculate the proportional fade length based on the note's duration. # At level 1.0, this is 10% of the note's start and 10% of its end. fade_percentage = 0.1 * params.s8bit_smooth_notes_level proportional_fade_samples = int(num_samples * fade_percentage) # 3. Calculate the absolute maximum fade length in samples. absolute_max_fade_samples = int(fs * max_fade_duration_s) # 4. The final fade_samples is the SMALLEST of the three constraints: # a) The proportional length. # b) The absolute maximum length. # c) Half the note's total length (to prevent overlap). fade_samples = min(proportional_fade_samples, absolute_max_fade_samples, num_samples // 2) if fade_samples > 0: # Apply a fade-in to the attack portion of the envelope. envelope[:fade_samples] *= np.linspace(0.5, 1.0, fade_samples) # Apply a fade-out to the tail portion of the envelope. envelope[-fade_samples:] *= np.linspace(1.0, 0.0, fade_samples) # Apply envelope to the (potentially combined) waveform note_waveform *= envelope # ========================================================================= # === Echo Sustain Logic for Long Plucky Notes (Now works correctly) === # ========================================================================= # This feature fills the silent tail of long notes with decaying echoes. # It is applied only for Plucky envelopes and after the main envelope has been applied. if params.s8bit_envelope_type == 'Plucky (AD Envelope)' and params.s8bit_echo_sustain and num_samples > 0: # The duration of the initial pluck is determined by its decay time. initial_pluck_duration_s = params.s8bit_decay_time_s initial_pluck_samples = int(initial_pluck_duration_s * fs) # Check if the note is long enough to even need echoes. if num_samples > initial_pluck_samples * params.s8bit_echo_trigger_threshold: # Only trigger if there's significant empty space. # Calculate the properties of the echoes. echo_delay_samples = int(fs / params.s8bit_echo_rate_hz) if echo_delay_samples > 0: # Prevent infinite loops echo_amplitude = start_amp * params.s8bit_echo_decay_factor # Start placing echoes after the first pluck has finished. current_sample_offset = initial_pluck_samples while current_sample_offset < num_samples: # Ensure there's space for a new echo. if current_sample_offset + echo_delay_samples <= num_samples: # Create a very short, plucky envelope for the echo. echo_attack_samples = min(int(0.002 * fs), echo_delay_samples) # 2ms attack echo_decay_samples = echo_delay_samples - echo_attack_samples if echo_decay_samples > 0: # Create the small echo envelope shape. echo_envelope = np.zeros(echo_delay_samples) echo_envelope[:echo_attack_samples] = np.linspace(0, echo_amplitude, echo_attack_samples) echo_envelope[echo_attack_samples:] = np.linspace(echo_amplitude, 0, echo_decay_samples) # Create a temporary waveform for the echo and apply the envelope. # It reuses the main note's frequency and oscillator phase. # Re-calculating phase here is simpler than tracking, for additive synthesis phase_inc_echo = 2 * np.pi * freq / fs phase_echo = np.cumsum(np.full(echo_delay_samples, phase_inc_echo)) if params.s8bit_waveform_type == 'Square': echo_waveform_segment = signal.square(phase_echo, duty=params.s8bit_pulse_width) elif params.s8bit_waveform_type == 'Sawtooth': echo_waveform_segment = signal.sawtooth(phase_echo) else: # Triangle echo_waveform_segment = signal.sawtooth(phase_echo, width=0.5) # Add the enveloped echo on top of the already-enveloped main waveform note_waveform[current_sample_offset : current_sample_offset + echo_delay_samples] += echo_waveform_segment * echo_envelope # Prepare for the next echo. echo_amplitude *= params.s8bit_echo_decay_factor current_sample_offset += echo_delay_samples # --- END of Echo Sustain Logic --- # --- Final Processing Stage (Per-Note) --- if use_aa: # 1. Frequency-dependent lowpass filter harm_limit = getattr(params, 's8bit_harmonic_lowpass_factor', 12.0) cutoff = min(fs * 0.45, max(3000.0, freq * harm_limit)) note_waveform = one_pole_lowpass(note_waveform, cutoff_hz=cutoff, fs=fs) # 2. Final Gain and Soft Limiter final_gain = getattr(params, 's8bit_final_gain', 0.8) note_waveform *= final_gain note_waveform = np.tanh(note_waveform) # Soft clip/limit # --- Add to main waveform buffer --- start_sample = int(note.start * fs) end_sample = start_sample + num_samples if end_sample > waveform.shape[1]: end_sample = waveform.shape[1] note_waveform = note_waveform[:end_sample-start_sample] # Add the mono note waveform to the stereo buffer with panning waveform[0, start_sample:end_sample] += note_waveform * pan_l waveform[1, start_sample:end_sample] += note_waveform * pan_r return waveform # Returns a (2, N) numpy array def analyze_midi_velocity(midi_path): midi = pretty_midi.PrettyMIDI(midi_path) all_velocities = [] print(f"Analyzing velocity for MIDI: {midi_path}") for i, instrument in enumerate(midi.instruments): velocities = [note.velocity for note in instrument.notes] all_velocities.extend(velocities) if velocities: print(f"Instrument {i} ({instrument.name}):") print(f" Notes count: {len(velocities)}") print(f" Velocity min: {min(velocities)}") print(f" Velocity max: {max(velocities)}") print(f" Velocity mean: {np.mean(velocities):.2f}") else: print(f"Instrument {i} ({instrument.name}): no notes found.") if all_velocities: print("\nOverall MIDI velocity stats:") print(f" Total notes: {len(all_velocities)}") print(f" Velocity min: {min(all_velocities)}") print(f" Velocity max: {max(all_velocities)}") print(f" Velocity mean: {np.mean(all_velocities):.2f}") else: print("No notes found in this MIDI.") def preview_sound_source(sound_source_name: str, *args): """ Generates a short audio preview for either a selected SoundFont or the 8-bit Synthesizer, using the Super Mario Bros. theme as a test melody. This function acts as a router: - If a SoundFont is selected, it uses FluidSynth. - If the 8-bit Synthesizer is selected, it uses the internal `synthesize_8bit_style` function, capturing the current UI settings for an accurate preview. Args: sound_source_name (str): The name of the SoundFont or the 8-bit synth label. *args: Captures all current UI settings, which are passed to build an AppParameters object for the 8-bit synth preview. Returns: A Gradio-compatible audio tuple (sample_rate, numpy_array). """ srate = 44100 # Use a standard sample rate for all previews. # 1. Create a MIDI object in memory. preview_midi = pretty_midi.PrettyMIDI() # Use a lead instrument. Program 81 (Lead 2, sawtooth) is a good, bright default. instrument = pretty_midi.Instrument(program=81, is_drum=False, name="Preview Lead") # 2. Define the melody: Super Mario Bros. theme intro # - tempo: A brisk 200 BPM, so each 0.15s step is a 16th note. # - notes: A list of tuples (pitch, duration_in_steps) tempo = 200.0 time_per_step = 60.0 / tempo / 2 # 16th note duration at this tempo # (Pitch, Duration in steps) # MIDI Pitch 60 = C4 (Middle C) melody_data = [ (76, 1), (76, 2), (76, 2), (72, 1), (76, 2), # E E E C E (79, 4), (67, 4) # G G(low) ] current_time = 0.0 for pitch, duration_steps in melody_data: start_time = current_time end_time = start_time + (duration_steps * time_per_step) # Add a tiny gap between notes to ensure they re-trigger clearly note_end_time = end_time - 0.01 note = pretty_midi.Note( velocity=120, # Use a high velocity for a bright, clear sound pitch=pitch, start=start_time, end=note_end_time ) instrument.notes.append(note) current_time = end_time preview_midi.instruments.append(instrument) # --- ROUTING LOGIC: Decide which synthesizer to use --- # CASE 1: 8-bit Synthesizer Preview if sound_source_name == SYNTH_8_BIT_LABEL: print("Generating preview for: 8-bit Synthesizer") try: # Create a temporary AppParameters object from the current UI settings params = AppParameters(**dict(zip(ALL_PARAM_KEYS, args))) # Use the internal synthesizer to render the preview MIDI audio_waveform = synthesize_8bit_style(midi_data=preview_midi, fs=srate, params=params) # Normalize and prepare for Gradio peak_val = np.max(np.abs(audio_waveform)) if peak_val > 0: audio_waveform /= peak_val # The synth returns (channels, samples), Gradio needs (samples, channels) audio_out = (audio_waveform.T * 32767).astype(np.int16) print("8-bit preview generated successfully.") return (srate, audio_out) except Exception as e: print(f"An error occurred during 8-bit preview generation: {e}") return None # CASE 2: SoundFont Preview else: soundfont_path = soundfonts_dict.get(sound_source_name) if not soundfont_path or not os.path.exists(soundfont_path): print(f"Preview failed: SoundFont file not found at '{soundfont_path}'") raise gr.Error(f"Could not find the SoundFont file for '{sound_source_name}'.") try: print(f"Generating preview for: {sound_source_name}") # Convert the in-memory MIDI object to a binary stream. midi_io = io.BytesIO() preview_midi.write(midi_io) midi_data = midi_io.getvalue() # Use the existing rendering function to generate the audio. # Ensure the output is a tuple (sample_rate, numpy_array) audio_out = midi_to_colab_audio( midi_data, soundfont_path=soundfont_path, sample_rate=srate, output_for_gradio=True ) # Ensure the returned value is exactly what Gradio expects. # The function `midi_to_colab_audio` should return a NumPy array. # We must wrap it in a tuple with the sample rate. if isinstance(audio_out, np.ndarray): print("SoundFont preview generated successfully.") return (srate, audio_out) else: # If the rendering function fails, it might return something else. # We handle this to prevent the Gradio error. print("Preview failed: Rendering function did not return valid audio data.") return None except Exception as e: # Catch any other errors, including from FluidSynth, and report them. print(f"An error occurred during SoundFont preview generation: {e}") # It's better to return None than to crash the UI. # The error will be visible in the console. return None def scale_instrument_velocity(instrument, scale=0.8): for note in instrument.notes: note.velocity = max(1, min(127, int(note.velocity * scale))) def normalize_loudness(audio_data, sample_rate, target_lufs=-23.0): """ Normalizes the audio data to a target integrated loudness (LUFS). This provides more consistent perceived volume than peak normalization. Args: audio_data (np.ndarray): The audio signal. sample_rate (int): The sample rate of the audio. target_lufs (float): The target loudness in LUFS. Defaults to -23.0, a common standard for broadcast. Returns: np.ndarray: The loudness-normalized audio data. """ try: # 1. Measure the integrated loudness of the input audio meter = pyln.Meter(sample_rate) # create meter loudness = meter.integrated_loudness(audio_data) # 2. Calculate the gain needed to reach the target loudness # The gain is applied in the linear domain, so we convert from dB loudness_gain_db = target_lufs - loudness loudness_gain_linear = 10.0 ** (loudness_gain_db / 20.0) # 3. Apply the gain normalized_audio = audio_data * loudness_gain_linear # 4. Final safety check: peak normalize to prevent clipping, just in case # the loudness normalization results in peaks > 1.0 peak_val = np.max(np.abs(normalized_audio)) if peak_val > 1.0: normalized_audio /= peak_val print(f"Warning: Loudness normalization resulted in clipping. Audio was peak-normalized as a safeguard.") print(f"Audio normalized from {loudness:.2f} LUFS to target {target_lufs} LUFS.") return normalized_audio except Exception as e: print(f"Loudness normalization failed: {e}. Falling back to original audio.") return audio_data # ================================================================================================= # === MIDI Merging Function === # ================================================================================================= def merge_midis(midi_path_left, midi_path_right, output_path): """ Merges two MIDI files into a single MIDI file. This robust version iterates through ALL instruments in both MIDI files, ensuring no data is lost if the source files are multi-instrumental. It applies hard-left panning (Pan=0) to every instrument from the left MIDI and hard-right panning (Pan=127) to every instrument from the right MIDI. """ try: analyze_midi_velocity(midi_path_left) analyze_midi_velocity(midi_path_right) midi_left = pretty_midi.PrettyMIDI(midi_path_left) midi_right = pretty_midi.PrettyMIDI(midi_path_right) merged_midi = pretty_midi.PrettyMIDI() # --- Process ALL instruments from the left channel MIDI --- if midi_left.instruments: print(f"Found {len(midi_left.instruments)} instrument(s) in the left channel MIDI.") # Use a loop to iterate through every instrument for instrument in midi_left.instruments: scale_instrument_velocity(instrument, scale=0.8) # To avoid confusion, we can prefix the instrument name instrument.name = f"Left - {instrument.name if instrument.name else 'Instrument'}" # Create and add the Pan Left control change # Create a Control Change event for Pan (controller number 10). # Set its value to 0 for hard left panning. # Add it at the very beginning of the track (time=0.0). pan_left = pretty_midi.ControlChange(number=10, value=0, time=0.0) # Use insert() to ensure the pan event is the very first one instrument.control_changes.insert(0, pan_left) # Append the fully processed instrument to the merged MIDI merged_midi.instruments.append(instrument) # --- Process ALL instruments from the right channel MIDI --- if midi_right.instruments: print(f"Found {len(midi_right.instruments)} instrument(s) in the right channel MIDI.") # Use a loop here as well for instrument in midi_right.instruments: scale_instrument_velocity(instrument, scale=0.8) instrument.name = f"Right - {instrument.name if instrument.name else 'Instrument'}" # Create and add the Pan Right control change # Create a Control Change event for Pan (controller number 10). # Set its value to 127 for hard right panning. # Add it at the very beginning of the track (time=0.0). pan_right = pretty_midi.ControlChange(number=10, value=127, time=0.0) instrument.control_changes.insert(0, pan_right) merged_midi.instruments.append(instrument) merged_midi.write(output_path) print(f"Successfully merged all instruments and panned into '{os.path.basename(output_path)}'") analyze_midi_velocity(output_path) return output_path except Exception as e: print(f"Error merging MIDI files: {e}") # Fallback logic remains the same if os.path.exists(midi_path_left): print("Fallback: Using only the left channel MIDI.") return midi_path_left return None def is_stereo_midi(midi_path: str) -> bool: """ Checks if a MIDI file contains the specific stereo panning control changes (hard left and hard right) created by the merge_midis function. Args: midi_path (str): The file path to the MIDI file. Returns: bool: True if both hard-left (0) and hard-right (127) pan controls are found, False otherwise. """ try: midi_data = pretty_midi.PrettyMIDI(midi_path) found_left_pan = False found_right_pan = False for instrument in midi_data.instruments: for control_change in instrument.control_changes: # MIDI Controller Number 10 is for Panning. if control_change.number == 10: if control_change.value == 0: found_left_pan = True elif control_change.value == 127: found_right_pan = True # Optimization: If we've already found both, no need to check further. if found_left_pan and found_right_pan: return True return found_left_pan and found_right_pan except Exception as e: # If the MIDI file is invalid or another error occurs, assume it's not our special stereo format. print(f"Could not analyze MIDI for stereo info: {e}") return False # ================================================================================================= # === Stage 1: Audio to MIDI Transcription Functions === # ================================================================================================= def TranscribePianoAudio(input_file): """ Transcribes a WAV or MP3 audio file of a SOLO PIANO performance into a MIDI file. This uses the ByteDance model. Args: input_file_path (str): The path to the input audio file. Returns: str: The file path of the generated MIDI file. """ print('=' * 70) print('STAGE 1: Starting Piano-Specific Transcription') print('=' * 70) # Generate a unique output filename for the MIDI fn = os.path.basename(input_file) fn1 = fn.split('.')[0] # Use os.path.join to create a platform-independent directory path output_dir = os.path.join("output", "transcribed_piano_") out_mid_path = os.path.join(output_dir, fn1 + '.mid') # Check for the directory's existence and create it if necessary if not os.path.exists(output_dir): os.makedirs(output_dir) print('-' * 70) print(f'Input file name: {fn}') print(f'Output MIDI path: {out_mid_path}') print('-' * 70) # Load audio using the utility function print('Loading audio...') (audio, _) = utilities.load_audio(input_file, sr=transcription_sample_rate, mono=True) print('Audio loaded successfully.') print('-' * 70) # Initialize the transcription model # Use 'cuda' if a GPU is available and configured, otherwise 'cpu' device = 'cuda' if torch.cuda.is_available() else 'cpu' print(f'Loading transcriptor model... device= {device}') transcriptor = PianoTranscription(device=device, checkpoint_path="src/models/CRNN_note_F1=0.9677_pedal_F1=0.9186.pth") print('Transcriptor loaded.') print('-' * 70) # Perform transcription print('Transcribing audio to MIDI (Piano-Specific)...') # This function call saves the MIDI file to the specified path transcriptor.transcribe(audio, out_mid_path) print('Piano transcription complete.') print('=' * 70) # Return the path to the newly created MIDI file return out_mid_path def TranscribeGeneralAudio(input_file, onset_threshold, frame_threshold, minimum_note_length, minimum_frequency, maximum_frequency, infer_onsets, melodia_trick, multiple_bends): """ Transcribes a general audio file into a MIDI file using basic-pitch. This is suitable for various instruments and vocals. """ print('=' * 70) print('STAGE 1: Starting General Purpose Transcription') print('=' * 70) fn = os.path.basename(input_file) fn1 = fn.split('.')[0] output_dir = os.path.join("output", "transcribed_general_") out_mid_path = os.path.join(output_dir, fn1 + '.mid') os.makedirs(output_dir, exist_ok=True) print(f'Input file: {fn}\nOutput MIDI: {out_mid_path}') # --- Perform transcription using basic-pitch --- print('Transcribing audio to MIDI (General Purpose)...') # The predict function handles audio loading internally model_output, midi_data, note_events = basic_pitch.inference.predict( audio_path=input_file, model_or_model_path=ICASSP_2022_MODEL_PATH, onset_threshold=onset_threshold, frame_threshold=frame_threshold, minimum_note_length=minimum_note_length, minimum_frequency=minimum_frequency, maximum_frequency=maximum_frequency, infer_onsets=infer_onsets, melodia_trick=melodia_trick, multiple_pitch_bends=multiple_bends ) # --- Save the MIDI file --- midi_data.write(out_mid_path) print('General transcription complete.') print('=' * 70) return out_mid_path # ================================================================================================= # === Stage 2: MIDI Transformation and Rendering Function === # ================================================================================================= def Render_MIDI(*, input_midi_path: str, params: AppParameters, progress: gr.Progress = None): """ Processes and renders a MIDI file according to user-defined settings. Can render using SoundFonts or a custom 8-bit synthesizer. Args: input_midi_path (str): The path to the input MIDI file. All other arguments are rendering options from the Gradio UI. Returns: A tuple containing all the output elements for the Gradio UI. """ print('*' * 70) print('STAGE 2: Starting MIDI Rendering') print('*' * 70) # --- File and Settings Setup --- fn = os.path.basename(input_midi_path) fn1 = fn.split('.')[0] # Use os.path.join to create a platform-independent directory path output_dir = os.path.join("output", "rendered_midi") if not os.path.exists(output_dir): os.makedirs(output_dir) # Now, join the clean directory path with the filename new_fn_path = os.path.join(output_dir, fn1 + '_rendered.mid') try: with open(input_midi_path, 'rb') as f: fdata = f.read() input_midi_md5hash = hashlib.md5(fdata).hexdigest() except FileNotFoundError: # Handle cases where the input file might not exist print(f"Error: Input MIDI file not found at {input_midi_path}") return [None] * 7 # Return empty values for all outputs print('=' * 70) print('Requested settings:') print(f'Input MIDI file name: {fn}') print(f'Input MIDI md5 hash: {input_midi_md5hash}') print('-' * 70) print(f"Render type: {params.render_type}") print(f"Soundfont bank: {params.soundfont_bank}") print(f"Audio render sample rate: {params.render_sample_rate}") # ... (add other print statements for settings if needed) print('=' * 70) # --- MIDI Processing using TMIDIX --- print('Processing MIDI... Please wait...') raw_score = MIDI.midi2single_track_ms_score(fdata) # call the function and store the returned list in a variable. processed_scores = TMIDIX.advanced_score_processor(raw_score, return_enhanced_score_notes=True, apply_sustain=params.render_with_sustains) # check if the returned list is empty. This happens when transcription finds no notes. # This check prevents the 'IndexError: list index out of range'. if not processed_scores: # If it is empty, print a warning and return a user-friendly error message. print("Warning: MIDI file contains no processable notes.") # The number of returned values must match the function's expected output. # We return a tuple with empty or placeholder values for all 7 output components. return ("N/A", fn1, "MIDI file contains no notes.", None, None, None, "No notes found.") # If the list is not empty, it is now safe to get the first element. escore = processed_scores[0] # Handle cases where the MIDI might not contain any notes if not escore: print("Warning: MIDI file contains no processable notes.") return ("N/A", fn1, "MIDI file contains no notes.",None, None, None, "No notes found.") # This line will now work correctly because merge_misaligned_notes is guaranteed to be an integer. if params.merge_misaligned_notes > 0: escore = TMIDIX.merge_escore_notes(escore, merge_threshold=params.merge_misaligned_notes) escore = TMIDIX.augment_enhanced_score_notes(escore, timings_divider=1) first_note_index = [e[0] for e in raw_score[1]].index('note') cscore = TMIDIX.chordify_score([1000, escore]) meta_data = raw_score[1][:first_note_index] + [escore[0]] + [escore[-1]] + [raw_score[1][-1]] aux_escore_notes = TMIDIX.augment_enhanced_score_notes(escore, sort_drums_last=True) song_description = TMIDIX.escore_notes_to_text_description(aux_escore_notes) print('Done!') print('=' * 70) print('Input MIDI metadata:', meta_data[:5]) print('=' * 70) print('Input MIDI song description:', song_description) print('=' * 70) print('Processing...Please wait...') # A deep copy of the score to be modified output_score = copy.deepcopy(escore) # Apply transformations based on render_type if params.render_type == "Extract melody": output_score = TMIDIX.add_melody_to_enhanced_score_notes(escore, return_melody=True) output_score = TMIDIX.recalculate_score_timings(output_score) elif params.render_type == "Flip": output_score = TMIDIX.flip_enhanced_score_notes(escore) elif params.render_type == "Reverse": output_score = TMIDIX.reverse_enhanced_score_notes(escore) elif params.render_type == 'Repair Durations': output_score = TMIDIX.fix_escore_notes_durations(escore, min_notes_gap=0) elif params.render_type == 'Repair Chords': fixed_cscore = TMIDIX.advanced_check_and_fix_chords_in_chordified_score(cscore)[0] output_score = TMIDIX.flatten(fixed_cscore) elif params.render_type == 'Remove Duplicate Pitches': output_score = TMIDIX.remove_duplicate_pitches_from_escore_notes(escore) elif params.render_type == "Add Drum Track": nd_escore = [e for e in escore if e[3] != 9] nd_escore = TMIDIX.augment_enhanced_score_notes(nd_escore) output_score = TMIDIX.advanced_add_drums_to_escore_notes(nd_escore) for e in output_score: e[1] *= 16 e[2] *= 16 print('MIDI processing complete.') print('=' * 70) # --- Final Processing and Patching --- if params.render_type != "Render as-is": print('Applying final adjustments (transpose, align, patch)...') if params.custom_render_patch != -1: # -1 indicates no change for e in output_score: if e[3] != 9: # not a drum channel e[6] = params.custom_render_patch if params.render_transpose_value != 0: output_score = TMIDIX.transpose_escore_notes(output_score, params.render_transpose_value) if params.render_transpose_to_C4: output_score = TMIDIX.transpose_escore_notes_to_pitch(output_score, 60) # C4 is MIDI pitch 60 if params.render_align == "Start Times": output_score = TMIDIX.recalculate_score_timings(output_score) output_score = TMIDIX.align_escore_notes_to_bars(output_score) elif params.render_align == "Start Times and Durations": output_score = TMIDIX.recalculate_score_timings(output_score) output_score = TMIDIX.align_escore_notes_to_bars(output_score, trim_durations=True) elif params.render_align == "Start Times and Split Durations": output_score = TMIDIX.recalculate_score_timings(output_score) output_score = TMIDIX.align_escore_notes_to_bars(output_score, split_durations=True) if params.render_type == "Longest Repeating Phrase": zscore = TMIDIX.recalculate_score_timings(output_score) lrno_score = TMIDIX.escore_notes_lrno_pattern_fast(zscore) if lrno_score is not None: output_score = lrno_score else: output_score = TMIDIX.recalculate_score_timings(TMIDIX.escore_notes_middle(output_score, 50)) if params.render_type == "Multi-Instrumental Summary": zscore = TMIDIX.recalculate_score_timings(output_score) c_escore_notes = TMIDIX.compress_patches_in_escore_notes_chords(zscore) if len(c_escore_notes) > 128: cmatrix = TMIDIX.escore_notes_to_image_matrix(c_escore_notes, filter_out_zero_rows=True, filter_out_duplicate_rows=True) smatrix = TPLOTS.square_image_matrix(cmatrix, num_pca_components=max(1, min(5, len(c_escore_notes) // 128))) output_score = TMIDIX.image_matrix_to_original_escore_notes(smatrix) for o in output_score: o[1] *= 250 o[2] *= 250 if params.render_output_as_solo_piano: output_score = TMIDIX.solo_piano_escore_notes(output_score, keep_drums=(not params.render_remove_drums)) if params.render_remove_drums and not params.render_output_as_solo_piano: output_score = TMIDIX.strip_drums_from_escore_notes(output_score) if params.render_type == "Solo Piano Summary": sp_escore_notes = TMIDIX.solo_piano_escore_notes(output_score, keep_drums=False) zscore = TMIDIX.recalculate_score_timings(sp_escore_notes) if len(zscore) > 128: bmatrix = TMIDIX.escore_notes_to_binary_matrix(zscore) cmatrix = TMIDIX.compress_binary_matrix(bmatrix, only_compress_zeros=True) smatrix = TPLOTS.square_binary_matrix(cmatrix, interpolation_order=max(1, min(5, len(zscore) // 128))) output_score = TMIDIX.binary_matrix_to_original_escore_notes(smatrix) for o in output_score: o[1] *= 200 o[2] *= 200 print('Final adjustments complete.') print('=' * 70) # --- Saving Processed MIDI File --- # Save the transformed MIDI data SONG, patches, _ = TMIDIX.patch_enhanced_score_notes(output_score) # The underlying function mistakenly adds a '.mid' extension. # We must pass the path without the extension to compensate. path_without_ext = new_fn_path.rsplit('.mid', 1)[0] MIDI.Tegridy_ms_SONG_to_MIDI_Converter(SONG, output_signature = 'Integrated-MIDI-Processor', output_file_name = path_without_ext, track_name='Processed Track', list_of_MIDI_patches=patches ) midi_to_render_path = new_fn_path else: # If "Render as-is", use the original MIDI data with open(new_fn_path, 'wb') as f: f.write(fdata) midi_to_render_path = new_fn_path # --- Audio Rendering --- print('Rendering final audio...') # Select sample rate srate = int(params.render_sample_rate) # --- Conditional Rendering Logic --- if params.soundfont_bank == SYNTH_8_BIT_LABEL: print("Using 8-bit style synthesizer...") try: # Load the MIDI file with pretty_midi for manual synthesis midi_data_for_synth = pretty_midi.PrettyMIDI(midi_to_render_path) # Synthesize the waveform # --- Passing new FX parameters to the synthesis function --- audio = synthesize_8bit_style(midi_data=midi_data_for_synth, fs=srate, params=params, progress=progress) # Normalize and prepare for Gradio peak_val = np.max(np.abs(audio)) if peak_val > 0: audio /= peak_val # Transpose from (2, N) to (N, 2) and convert to int16 for Gradio audio_out = (audio.T * 32767).astype(np.int16) except Exception as e: print(f"Error during 8-bit synthesis: {e}") return [None] * 7 else: print(f"Using SoundFont: {params.soundfont_bank}") # Get the full path from the global dictionary soundfont_path = soundfonts_dict.get(params.soundfont_bank) # Select soundfont if not soundfont_path or not os.path.exists(soundfont_path): # If the selected soundfont is not found, inform the user directly via the UI. raise gr.Error(f"SoundFont file '{params.soundfont_bank}' could not be found. Please check your 'src/sf2' directory or select another SoundFont.") # # Error handling in case the selected file is not found # error_msg = f"SoundFont '{params.soundfont_bank}' not found!" # print(f"ERROR: {error_msg}") # # Fallback to the first available soundfont if possible # if soundfonts_dict: # fallback_key = list(soundfonts_dict.keys())[0] # soundfont_path = soundfonts_dict[fallback_key] # print(f"Falling back to '{fallback_key}'.") # else: # # If no soundfonts are available at all, raise an error # raise gr.Error("No SoundFonts are available for rendering!") with open(midi_to_render_path, 'rb') as f: midi_file_content = f.read() audio_out = midi_to_colab_audio(midi_file_content, soundfont_path=soundfont_path, # Use the dynamically found path sample_rate=srate, output_for_gradio=True ) print('Audio rendering complete.') print('=' * 70) # --- Preparing Outputs for Gradio --- with open(midi_to_render_path, 'rb') as f: new_md5_hash = hashlib.md5(f.read()).hexdigest() output_plot = TPLOTS.plot_ms_SONG(output_score, plot_title=f"Score of {fn1}", return_plt=True) output_midi_summary = str(meta_data) return new_md5_hash, fn1, output_midi_summary, midi_to_render_path, (srate, audio_out), output_plot, song_description def analyze_midi_features(midi_data): """ Analyzes a PrettyMIDI object to extract musical features for parameter recommendation. Args: midi_data (pretty_midi.PrettyMIDI): The MIDI data to analyze. Returns: dict or None: A dictionary containing features, or None if the MIDI is empty. Features: 'note_count', 'instruments_count', 'duration', 'note_density', 'avg_velocity', 'pitch_range'. """ all_notes = [note for instrument in midi_data.instruments for note in instrument.notes] note_count = len(all_notes) # Return None if the MIDI file has no notes to analyze. if note_count == 0: return None duration = midi_data.get_end_time() # Avoid division by zero for empty-duration MIDI files. if duration == 0: note_density = 0 else: note_density = note_count / duration # --- Calculate new required features --- avg_velocity = sum(note.velocity for note in all_notes) / note_count avg_pitch = sum(note.pitch for note in all_notes) / note_count avg_note_length = sum(note.end - note.start for note in all_notes) / note_count # Calculate pitch range if note_count > 1: min_pitch = min(note.pitch for note in all_notes) max_pitch = max(note.pitch for note in all_notes) pitch_range = max_pitch - min_pitch else: pitch_range = 0 return { 'note_count': note_count, 'instruments_count': len(midi_data.instruments), 'duration': duration, 'note_density': note_density, # Notes per second 'avg_velocity': avg_velocity, 'pitch_range': pitch_range, # In semitones 'avg_pitch': avg_pitch, 'avg_note_length': avg_note_length, } def determine_waveform_type(features): """ Determines the best waveform type based on analyzed MIDI features. - Square: Best for most general-purpose, bright melodies. - Sawtooth: Best for intense, heavy, or powerful leads and basses. - Triangle: Best for soft, gentle basses or flute-like sounds. Args: features (dict): The dictionary of features from analyze_midi_features. Returns: str: The recommended waveform type ('Square', 'Sawtooth', or 'Triangle'). """ # 1. Check for conditions that strongly suggest a Triangle wave (soft bassline) # MIDI Pitch 52 is ~G#3. If the average pitch is below this, it's likely a bass part. # If notes are long and the pitch range is narrow, it confirms a simple, melodic bassline. if features['avg_pitch'] <= 52 and features['avg_note_length'] >= 0.3 and features['pitch_range'] < 12: return "Triangle" # 2. Check for conditions that suggest a Sawtooth wave (intense/complex part) # High note density or a very wide pitch range often indicates an aggressive lead or a complex solo. # The sawtooth's rich harmonics are perfect for this. if features['note_density'] >= 6 or features['pitch_range'] >= 18: return "Sawtooth" # 3. Default to the most versatile waveform: Square return "Square" def recommend_8bit_params(midi_data, default_preset): """ Recommends 8-bit synthesizer parameters using a unified, factor-based model. This "AI" generates a sound profile based on normalized musical features. Args: midi_data (pretty_midi.PrettyMIDI): The MIDI data to analyze. default_preset (dict): A fallback preset if analysis fails. Returns: dict: A dictionary of recommended synthesizer parameters. """ features = analyze_midi_features(midi_data) if features is None: # Return a default preset if MIDI is empty or cannot be analyzed return default_preset # --- Rule-based Parameter Recommendation --- params = {} # --- 1. Core Timbre Selection --- # Intelligent Waveform Selection params['waveform_type'] = determine_waveform_type(features) # Determine pulse width *after* knowing the waveform. # This only applies if the waveform is Square. if params['waveform_type'] == 'Square': # For Square waves, use pitch complexity to decide pulse width. # Complex melodies get a thinner sound (0.3) for clarity. # Simpler melodies get a fuller sound (0.5). params['pulse_width'] = 0.3 if features['pitch_range'] > 30 else 0.5 else: # For Sawtooth or Triangle, pulse width is not applicable. Set a default. params['pulse_width'] = 0.5 # --- 2. Envelope and Rhythm --- # Determine envelope type based on note density is_plucky = features['note_density'] > 10 params['envelope_type'] = 'Plucky (AD Envelope)' if is_plucky else 'Sustained (Full Decay)' params['decay_time_s'] = 0.15 if is_plucky else 0.4 # --- 3. Modulation (Vibrato) --- # Vibrato depth and rate based on velocity and density params['vibrato_depth'] = min(max((features['avg_velocity'] - 60) / 20, 0), 10) # More velocity = more depth if features['note_density'] > 12: params['vibrato_rate'] = 7.0 # Very fast music -> frantic vibrato elif features['note_density'] > 6: params['vibrato_rate'] = 5.0 # Moderately fast music -> standard vibrato else: params['vibrato_rate'] = 3.0 # Slow music -> gentle vibrato # --- 4. Progressive/Graded Parameters using Normalization --- # Smooth notes level (0.0 to 1.0): More smoothing for denser passages. # Effective range: 3 to 8 notes/sec. params['smooth_notes_level'] = min(max((features['note_density'] - 3) / 5.0, 0.0), 1.0) # Smoothen notes in denser passages # Continuous vibrato level (0.0 to 1.0): Less dense passages get more lyrical, continuous vibrato. # Effective range: 5 to 10 notes/sec. (Inverted) params['continuous_vibrato_level'] = 1.0 - min(max((features['note_density'] - 5) / 5.0, 0.0), 1.0) # Lyrical (less dense) music gets connected vibrato # Noise level (0.0 to 0.1): Higher velocity passages get more "air" or "grit". # Effective range: velocity 50 to 90. params['noise_level'] = min(max((features['avg_velocity'] - 50) / 40.0, 0.0), 1.0) * 0.1 # Distortion level (0.0 to 0.1): Shorter notes get more distortion for punch. # Effective range: note length 0.5s down to 0.25s. (Inverted) if features['avg_note_length'] < 0.25: # Short, staccato notes params['distortion_level'] = 0.1 elif features['avg_note_length'] < 0.5: # Medium length notes params['distortion_level'] = 0.05 else: # Long, sustained notes params['distortion_level'] = 0.0 # Progressive FM modulation based on a combined complexity factor. # Normalizes note density and pitch range to a 0-1 scale. density_factor = min(max((features['note_density'] - 5) / 15, 0), 1) # Effective range 5-20 notes/sec range_factor = min(max((features['pitch_range'] - 15) / 30, 0), 1) # Effective range 15-45 semitones # The overall complexity is the average of these two factors. complexity_factor = (density_factor + range_factor) / 2 params['fm_modulation_depth'] = round(0.3 * complexity_factor, 3) params['fm_modulation_rate'] = round(200 * complexity_factor, 1) # Non-linear bass boost # REFINED LOGIC: Non-linear bass boost based on instrument count. # More instruments lead to less bass boost to avoid a muddy mix, # while solo or duo arrangements get a significant boost to sound fuller. # The boost level has a floor of 0.2 and a ceiling of 1.0. params['bass_boost_level'] = max(0.2, 1.0 - (features['instruments_count'] - 1) * 0.15) # Round all float values for cleaner output for key, value in params.items(): if isinstance(value, float): params[key] = round(value, 3) return params # ================================================================================================= # === Main Application Logic === # ================================================================================================= # --- Helper function to encapsulate the transcription pipeline for a single audio file --- def _transcribe_stem(audio_path: str, base_name: str, temp_dir: str, params: AppParameters): """ Takes a single audio file path and runs the full transcription pipeline on it. This includes stereo/mono handling and normalization. Returns the file path of the resulting transcribed MIDI. """ print(f"\n--- Transcribing Stem: {os.path.basename(audio_path)} ---") # Load the audio stem to process it audio_data, native_sample_rate = librosa.load(audio_path, sr=None, mono=False) if params.enable_stereo_processing and audio_data.ndim == 2 and audio_data.shape[0] == 2: print("Stereo processing enabled for stem.") left_channel_np = audio_data[0] right_channel_np = audio_data[1] normalized_left = normalize_loudness(left_channel_np, native_sample_rate) normalized_right = normalize_loudness(right_channel_np, native_sample_rate) temp_left_path = os.path.join(temp_dir, f"{base_name}_left.flac") temp_right_path = os.path.join(temp_dir, f"{base_name}_right.flac") sf.write(temp_left_path, normalized_left, native_sample_rate) sf.write(temp_right_path, normalized_right, native_sample_rate) print(f"Saved left channel to: {temp_left_path}") print(f"Saved right channel to: {temp_right_path}") print("Transcribing left and right channel...") if params.transcription_method == "General Purpose": midi_path_left = TranscribeGeneralAudio(temp_left_path, params.onset_threshold, params.frame_threshold, params.minimum_note_length, params.minimum_frequency, params.maximum_frequency, params.infer_onsets, params.melodia_trick, params.multiple_pitch_bends) midi_path_right = TranscribeGeneralAudio(temp_right_path, params.onset_threshold, params.frame_threshold, params.minimum_note_length, params.minimum_frequency, params.maximum_frequency, params.infer_onsets, params.melodia_trick, params.multiple_pitch_bends) else: # Piano-Specific midi_path_left = TranscribePianoAudio(temp_left_path) midi_path_right = TranscribePianoAudio(temp_right_path) if midi_path_left and midi_path_right: merged_midi_path = os.path.join(temp_dir, f"{base_name}_merged.mid") return merge_midis(midi_path_left, midi_path_right, merged_midi_path) elif midi_path_left: print("Warning: Right channel transcription failed. Using left channel only.") return midi_path_left elif midi_path_right: print("Warning: Left channel transcription failed. Using right channel only.") return midi_path_right else: print(f"Warning: Stereo transcription failed for stem {base_name}.") return None else: print("Mono processing for stem.") mono_signal_np = np.mean(audio_data, axis=0) if audio_data.ndim > 1 else audio_data normalized_mono = normalize_loudness(mono_signal_np, native_sample_rate) temp_mono_path = os.path.join(temp_dir, f"{base_name}_mono.flac") sf.write(temp_mono_path, normalized_mono, native_sample_rate) if params.transcription_method == "General Purpose": return TranscribeGeneralAudio(temp_mono_path, params.onset_threshold, params.frame_threshold, params.minimum_note_length, params.minimum_frequency, params.maximum_frequency, params.infer_onsets, params.melodia_trick, params.multiple_pitch_bends) else: return TranscribePianoAudio(temp_mono_path) # --- The core processing engine for a single file --- def run_single_file_pipeline(input_file_path: str, timestamp: str, params: AppParameters, progress: gr.Progress = None): """ This is the main processing engine. It takes a file path and a dictionary of all settings, and performs the full pipeline: load, separate, transcribe, render, re-merge. It is UI-agnostic and returns file paths and data, not Gradio updates. It now accepts a Gradio Progress object to report granular progress. """ # Helper function to safely update progress def update_progress(fraction, desc): if progress: progress(fraction, desc=desc) # --- Start timer for this specific file --- file_start_time = reqtime.time() filename = os.path.basename(input_file_path) base_name = os.path.splitext(filename)[0] # --- Determine file type to select the correct progress timeline --- is_midi_input = filename.lower().endswith(('.mid', '.midi', '.kar')) update_progress(0, f"Starting: {filename}") print(f"\n{'='*20} Starting Pipeline for: {filename} {'='*20}") # --- Use the provided timestamp for unique filenames --- timestamped_base_name = f"{base_name}_{timestamp}" # --- Step 1: Check file type and transcribe if necessary --- if is_midi_input: # For MIDI files, we start at 0% and directly proceed to the rendering steps. update_progress(0, "MIDI file detected, skipping transcription...") print("MIDI file detected. Skipping transcription. Proceeding directly to rendering.") if is_stereo_midi(input_file_path): print("\nINFO: Stereo pan information (Left/Right) detected in the input MIDI. It will be rendered in stereo.\n") midi_path_for_rendering = input_file_path else: temp_dir = "output/temp_transcribe" # Define temp_dir early for the fallback os.makedirs(temp_dir, exist_ok=True) # --- Audio Loading --- update_progress(0.1, "Audio file detected, loading...") print("Audio file detected. Starting pre-processing...") # --- Robust audio loading with ffmpeg fallback --- try: # Try loading directly with torchaudio (efficient for supported formats). # This works for formats like WAV, MP3, FLAC, OGG, etc. print("Attempting to load audio with torchaudio...") audio_tensor, native_sample_rate = torchaudio.load(input_file_path) print("Torchaudio loading successful.") except Exception as e: update_progress(0.15, "Torchaudio failed, trying ffmpeg...") print(f"Torchaudio failed: {e}. Attempting fallback with ffmpeg...") try: # Define a path for the temporary converted file converted_flac_path = os.path.join(temp_dir, f"{timestamped_base_name}_converted.flac") # Use ffmpeg to convert the input file to a clean FLAC file on disk ( ffmpeg .input(input_file_path) .output(converted_flac_path, acodec='flac') .overwrite_output() .run(capture_stdout=True, capture_stderr=True) ) # Now, load the newly created and guaranteed-to-be-compatible FLAC file audio_tensor, native_sample_rate = torchaudio.load(converted_flac_path) print(f"FFmpeg fallback successful. Loaded from: {converted_flac_path}") except Exception as ffmpeg_err: # In batch mode, we just print an error and skip this file stderr = ffmpeg_err.stderr.decode() if hasattr(ffmpeg_err, 'stderr') else str(ffmpeg_err) print(f"ERROR: Could not load {filename}. Skipping. FFmpeg error: {stderr}") return None # Return None to indicate failure # --- Demucs Vocal Separation Logic --- # This block now handles multi-stem separation, transcription, and merging logic. separated_stems = {} # This will store the audio tensors for merging if params.separate_vocals and demucs_model is not None: # --- Vocal Separation Workflow --- update_progress(0.2, "Separating audio with Demucs...") # Convert to the format Demucs expects (e.g., 44.1kHz, stereo) audio_tensor = convert_audio(audio_tensor, native_sample_rate, demucs_model.samplerate, demucs_model.audio_channels) # Move tensor to GPU if available for faster processing if torch.cuda.is_available(): audio_tensor = audio_tensor.cuda() print("Separating audio with Demucs... This may take some time.") # --- Wrap the model call in a no_grad() context --- with torch.no_grad(): all_stems = apply_model( demucs_model, audio_tensor[None], # The input shape is [batch, channels, samples] device='cuda' if torch.cuda.is_available() else 'cpu', progress=True )[0] # Remove the batch dimension from the output # --- Clear CUDA cache immediately after use --- if torch.cuda.is_available(): torch.cuda.empty_cache() print("CUDA cache cleared.") sources = {name: stem for name, stem in zip(demucs_model.sources, all_stems)} # --- Store original stems for potential re-merging --- for name, tensor in sources.items(): separated_stems[name] = (tensor.cpu(), demucs_model.samplerate) # --- Prepare Stems for Transcription --- stems_to_transcribe = {} if params.enable_advanced_separation: # User is in advanced mode, handle each stem individually if params.transcribe_vocals: stems_to_transcribe['vocals'] = sources['vocals'] if params.transcribe_drums: stems_to_transcribe['drums'] = sources['drums'] if params.transcribe_bass: stems_to_transcribe['bass'] = sources['bass'] if params.transcribe_other_or_accompaniment: stems_to_transcribe['other'] = sources['other'] else: # User is in simple mode, create a single 'accompaniment' stem accompaniment_tensor = sources['drums'] + sources['bass'] + sources['other'] if params.transcribe_vocals: stems_to_transcribe['vocals'] = sources['vocals'] if params.transcribe_other_or_accompaniment: stems_to_transcribe['accompaniment'] = accompaniment_tensor # --- Transcribe Selected Stems --- transcribed_midi_paths = [] if stems_to_transcribe: stem_count = len(stems_to_transcribe) for i, (name, tensor) in enumerate(stems_to_transcribe.items()): update_progress(0.3 + (0.3 * (i / stem_count)), f"Transcribing stem: {name}...") stem_path = os.path.join(temp_dir, f"{timestamped_base_name}_{name}.flac") torchaudio.save(stem_path, tensor.cpu(), demucs_model.samplerate) midi_path = _transcribe_stem(stem_path, f"{timestamped_base_name}_{name}", temp_dir, params) if midi_path: transcribed_midi_paths.append((name, midi_path)) # --- Merge Transcribed MIDIs --- if not transcribed_midi_paths: raise gr.Error("Separation was enabled, but no stems were selected for transcription, or transcription failed.") elif len(transcribed_midi_paths) == 1: midi_path_for_rendering = transcribed_midi_paths[0][1] else: update_progress(0.6, "Merging transcribed MIDIs...") merged_midi = pretty_midi.PrettyMIDI() for name, path in transcribed_midi_paths: try: midi_stem = pretty_midi.PrettyMIDI(path) for inst in midi_stem.instruments: inst.name = f"{name.capitalize()} - {inst.name}" merged_midi.instruments.append(inst) except Exception as e: print(f"Warning: Could not merge MIDI for stem {name}. Error: {e}") final_merged_midi_path = os.path.join(temp_dir, f"{timestamped_base_name}_full_transcription.mid") merged_midi.write(final_merged_midi_path) midi_path_for_rendering = final_merged_midi_path else: # Standard workflow without separation # --- Standard Workflow: Transcribe the original full audio --- audio_to_transcribe_path = os.path.join(temp_dir, f"{timestamped_base_name}_original.flac") torchaudio.save(audio_to_transcribe_path, audio_tensor, native_sample_rate) update_progress(0.2, "Transcribing audio to MIDI...") midi_path_for_rendering = _transcribe_stem(audio_to_transcribe_path, f"{timestamped_base_name}_original", temp_dir, params) if not midi_path_for_rendering or not os.path.exists(midi_path_for_rendering): print(f"ERROR: Transcription failed for {filename}. Skipping.") return None # --- Step 2: Render the FINAL MIDI file with selected options --- # The progress values are now conditional based on the input file type. update_progress(0.1 if is_midi_input else 0.6, "Applying MIDI transformations...") # --- Auto-Recommendation Logic --- # If the user selected the auto-recommend option, override the parameters if params.s8bit_preset_selector == "Auto-Recommend (Analyze MIDI)": update_progress(0.15 if is_midi_input else 0.65, "Auto-recommending 8-bit parameters...") print("Auto-Recommendation is enabled. Analyzing MIDI features...") try: midi_to_analyze = pretty_midi.PrettyMIDI(midi_path_for_rendering) default_preset = S8BIT_PRESETS[FALLBACK_PRESET_NAME] recommended_params = recommend_8bit_params(midi_to_analyze, default_preset) print("Recommended parameters:", recommended_params) # Update the params object *before* the main pipeline runs for key, value in recommended_params.items(): setattr(params, f"s8bit_{key}", value) print("Parameters updated with recommendations.") except Exception as e: print(f"Could not auto-recommend parameters for {filename}: {e}.") # --- Step 2: Render the FINAL MIDI file --- update_progress(0.2 if is_midi_input else 0.7, "Rendering MIDI to audio...") print(f"Proceeding to render MIDI file: {os.path.basename(midi_path_for_rendering)}") # Call the rendering function, Pass dictionaries directly to Render_MIDI results_tuple = Render_MIDI(input_midi_path=midi_path_for_rendering, params=params, progress=progress) # --- Final Audio Merging Logic --- stems_to_merge = [] if params.separate_vocals: if params.merge_vocals_to_render and 'vocals' in separated_stems: stems_to_merge.append(separated_stems['vocals']) if params.enable_advanced_separation: if params.merge_drums_to_render and 'drums' in separated_stems: stems_to_merge.append(separated_stems['drums']) if params.merge_bass_to_render and 'bass' in separated_stems: stems_to_merge.append(separated_stems['bass']) if params.merge_other_or_accompaniment and 'other' in separated_stems: stems_to_merge.append(separated_stems['other']) else: # Simple mode if params.merge_other_or_accompaniment: # 'other' checkbox now controls the whole accompaniment accompaniment_tensor = separated_stems['drums'][0] + separated_stems['bass'][0] + separated_stems['other'][0] stems_to_merge.append((accompaniment_tensor, demucs_model.samplerate)) if stems_to_merge: update_progress(0.9, "Re-merging audio stems...") rendered_srate, rendered_music_int16 = results_tuple[4] rendered_music_float = rendered_music_int16.astype(np.float32) / 32767.0 final_mix_tensor = torch.from_numpy(rendered_music_float).T final_srate = rendered_srate for stem_tensor, stem_srate in stems_to_merge: # Resample if necessary if stem_srate != final_srate: # Resample all stems to match the rendered audio's sample rate resampler = torchaudio.transforms.Resample(stem_srate, final_srate) stem_tensor = resampler(stem_tensor) # Pad and add to the final mix len_mix = final_mix_tensor.shape[1] len_stem = stem_tensor.shape[1] if len_mix > len_stem: stem_tensor = torch.nn.functional.pad(stem_tensor, (0, len_mix - len_stem)) elif len_stem > len_mix: final_mix_tensor = torch.nn.functional.pad(final_mix_tensor, (0, len_stem - len_mix)) final_mix_tensor += stem_tensor # Normalize final mix to prevent clipping max_abs = torch.max(torch.abs(final_mix_tensor)) if max_abs > 1.0: final_mix_tensor /= max_abs # Convert back to the required format (int16 numpy array) merged_audio_int16 = (final_mix_tensor.T.numpy() * 32767).astype(np.int16) # Update the results tuple with the newly merged audio temp_results_list = list(results_tuple) temp_results_list[4] = (final_srate, merged_audio_int16) results_tuple = tuple(temp_results_list) # results_tuple is now updated print("Re-merging complete.") # --- Save final audio and return path --- update_progress(0.95, "Saving final files...") final_srate, final_audio_data = results_tuple[4] final_midi_path_from_render = results_tuple[3] # Get the path of the processed MIDI # --- Use timestamped names for final outputs --- output_audio_dir = "output/final_audio" output_midi_dir = "output/final_midi" os.makedirs(output_audio_dir, exist_ok=True) os.makedirs(output_midi_dir, exist_ok=True) final_audio_path = os.path.join(output_audio_dir, f"{timestamped_base_name}_rendered.flac") # Also, copy the final processed MIDI to a consistent output directory with a timestamped name final_midi_path = os.path.join(output_midi_dir, f"{timestamped_base_name}_processed.mid") sf.write(final_audio_path, final_audio_data, final_srate) # Use shutil to copy the final midi file to its new home shutil.copy(final_midi_path_from_render, final_midi_path) # --- Log the processing time for this specific file at the end --- file_processing_time = reqtime.time() - file_start_time print(f"--- Pipeline finished for {filename} in {file_processing_time:.2f} seconds. ---") print(f"Output Audio: {final_audio_path}\nOutput MIDI: {final_midi_path}") # Return a dictionary of all results for the wrappers to use results = { "final_audio_path": final_audio_path, "final_midi_path": final_midi_path, "md5_hash": results_tuple[0], "title": results_tuple[1], "summary": results_tuple[2], "plot": results_tuple[5], "description": results_tuple[6] } update_progress(1.0, "Done!") # Return both the results and the final state of the parameters object return results, params # ================================================================================================= # === Gradio UI Wrappers === # ================================================================================================= class BatchProgressTracker: """ A custom progress tracker for batch processing that can update a main progress bar and also create its own tqdm-style sub-progress bars. """ def __init__(self, main_progress: gr.Progress, total_files: int, current_file_index: int, filename: str): self._main_progress = main_progress self._total_files = total_files self._current_file_index = current_file_index self._filename = filename self._progress_per_file = 1 / total_files if total_files > 0 else 0 def __call__(self, local_fraction: float, desc: str = ""): """Makes the object callable like a function for simple progress updates.""" overall_fraction = (self._current_file_index / self._total_files) + (local_fraction * self._progress_per_file) full_desc = f"({self._current_file_index + 1}/{self._total_files}) {self._filename}: {desc}" # Update the main progress bar self._main_progress(overall_fraction, desc=full_desc) def tqdm(self, iterable, desc="", total=None): """Provides a tqdm method that delegates to the original gr.Progress object.""" # The description for the sub-progress bar tqdm_desc = f"({self._current_file_index + 1}/{self._total_files}) {self._filename}: {desc}" # Use the original gr.Progress object to create the tqdm iterator return self._main_progress.tqdm(iterable, desc=tqdm_desc, total=total) # --- Thin wrapper for batch processing --- def batch_process_files(input_files, progress=gr.Progress(track_tqdm=True), *args): """ Gradio wrapper for batch processing. It iterates through files, calls the core pipeline, and collects the output file paths. It now provides detailed, nested progress updates. """ if not input_files: print("No files uploaded for batch processing.") return [], [] # Return two empty lists # --- Start timer for the entire batch --- batch_start_time = reqtime.time() # --- Generate a single timestamp for the entire batch job --- batch_timestamp = reqtime.strftime("%Y%m%d-%H%M%S") # Create the AppParameters object from the flat list of UI values params = AppParameters(**dict(zip(ALL_PARAM_KEYS, args))) output_audio_paths = [] output_midi_paths = [] # List to collect MIDI file paths total_files = len(input_files) # Initialize progress at 0% progress(0, desc="Starting Batch Process...") for i, file_obj in enumerate(input_files): # The input from gr.File is a tempfile object, we need its path input_path = file_obj.name filename = os.path.basename(input_path) # --- Use the new BatchProgressTracker class --- # Instead of a simple function, create an instance of our tracker class. # This object can both update the main progress and has a .tqdm method. batch_progress_tracker = BatchProgressTracker( main_progress=progress, total_files=total_files, current_file_index=i, filename=filename ) # --- Pass the batch_timestamp to the pipeline --- results, _ = run_single_file_pipeline(input_path, batch_timestamp, copy.copy(params), progress=batch_progress_tracker) if results: if results.get("final_audio_path"): output_audio_paths.append(results["final_audio_path"]) if results.get("final_midi_path"): output_midi_paths.append(results["final_midi_path"]) # Collect MIDI path # Ensure the progress bar reaches 100% upon completion progress(1, desc="Batch Process Complete!") # --- Calculate and print the total batch time --- total_batch_time = reqtime.time() - batch_start_time print(f"\nBatch processing complete. {len(output_audio_paths)} of {total_files} files processed successfully.") print(f"Total batch execution time: {total_batch_time:.2f} seconds.") # --- Return both lists of paths --- return output_audio_paths, output_midi_paths # --- The original function is now a thin wrapper for the single file UI --- def process_and_render_file(input_file, *args, progress=gr.Progress()): """ Gradio wrapper for the single file processing UI. Packs UI values into an AppParameters object. Calls the core pipeline and formats the output for all UI components. Main function to handle file processing. It determines the file type and calls the appropriate functions for transcription and/or rendering based on user selections. Now includes a progress bar. """ if input_file is None: # Return a list of updates to clear all output fields and UI controls return [gr.update(value=None)] * (7 + 14) # 7 results + 14 UI controls (13 synth + 1 preset selector) # --- Start timer for the single file job --- job_start_time = reqtime.time() # --- Generate a timestamp for this single job --- single_file_timestamp = reqtime.strftime("%Y%m%d-%H%M%S") # Create the AppParameters object from the flat list of UI values # The first value in *args is s8bit_preset_selector, the rest match the keys params = AppParameters(input_file=input_file, **dict(zip(ALL_PARAM_KEYS, args))) # Run the core pipeline, passing the timestamp and progress to the pipeline results, final_params = run_single_file_pipeline(input_file, single_file_timestamp, params, progress=progress) if results is None: raise gr.Error("File processing failed. Check console for details.") # --- Calculate and print the total job time --- total_job_time = reqtime.time() - job_start_time print(f"Total single-file job execution time: {total_job_time:.2f} seconds.") # --- Prepare UI updates using the returned final_params --- # This ensures the UI always reflects the parameters that were actually used for the render. final_ui_updates = [] # Logic to decide what the preset selector should show after the run if params.s8bit_preset_selector == "Auto-Recommend (Analyze MIDI)": # After auto-recommendation, the state becomes "Custom" final_ui_updates.append(gr.update(value="Custom")) else: # Otherwise, just keep the user's current selection final_ui_updates.append(gr.update(value=final_params.s8bit_preset_selector)) # Get the keys for the 13 synthesizer controls (excluding the preset selector itself) s8bit_control_keys = [key for key in ALL_PARAM_KEYS if key.startswith('s8bit_') and key != 's8bit_preset_selector'] # Always update all 13 controls to match the final parameters used in the backend for key in s8bit_control_keys: value = getattr(final_params, key) # Explicitly convert numpy numeric types to native Python types. # This prevents them from being serialized as strings in the UI, # which would cause a TypeError on the next run. if isinstance(value, np.generic): value = value.item() final_ui_updates.append(value) # Format the main results for the output components main_results = [ results['md5_hash'], results['title'], results['summary'], results['final_midi_path'], results['final_audio_path'], results['plot'], results['description'] ] # The total return list now has a consistent structure and logic return main_results + final_ui_updates # ================================================================================================= # === Gradio UI Setup === # ================================================================================================= if __name__ == "__main__": # Initialize the app: download model (if needed) and apply patches # Set to False if you don't have 'requests' or 'tqdm' installed initialize_app() # --- Prepare soundfonts and make the map globally accessible --- global soundfonts_dict, demucs_model # On application start, download SoundFonts from Hugging Face Hub if they don't exist. soundfonts_dict = prepare_soundfonts() print(f"Found {len(soundfonts_dict)} local SoundFonts.") if not soundfonts_dict: print("\nWARNING: No SoundFonts were found or could be downloaded.") print("Rendering with SoundFonts will fail. Only the 8-bit synthesizer will be available.") # --- Pre-load the Demucs model on startup for efficiency --- print("Loading Demucs model (htdemucs_ft), this may take a moment on first run...") try: demucs_model = get_model(name='htdemucs_ft') if torch.cuda.is_available(): demucs_model = demucs_model.cuda() print("Demucs model loaded successfully.") except Exception as e: print(f"Warning: Could not load Demucs model. Vocal separation will not be available. Error: {e}") demucs_model = None # --- Dictionary containing descriptions for each render type --- RENDER_TYPE_DESCRIPTIONS = { "Render as-is": "**Mode: Pass-through.** Renders the MIDI file directly without any modifications. Advanced MIDI options will be ignored.", "Custom render": "**Mode: Activate Advanced Options.** Applies all settings from the 'Advanced MIDI Rendering Options' accordion without making other structural changes to the MIDI.", "Extract melody": "**Action: Simplify.** Analyzes all tracks and attempts to isolate and render only the main melody line.", "Flip": "**Action: Experimental.** Inverts the pitch of each note around the song's average pitch.", "Reverse": "**Action: Experimental.** Reverses the playback order of all notes in the MIDI file.", "Repair Durations": "**Action: Fix.** Recalculates note durations to ensure they connect smoothly (legato), filling any small gaps.", "Repair Chords": "**Action: Fix.** Analyzes and aligns notes that occur at similar times to form cleaner, more structured chords.", "Remove Duplicate Pitches": "**Action: Simplify.** If multiple instruments play the exact same pitch at the same time, it keeps only one.", "Longest Repeating Phrase": "**Action: Analyze.** Finds the longest, most-repeated musical phrase (often the chorus) and renders only that section.", "Multi-Instrumental Summary": "**Action: AI Summary.** Creates a short, compressed summary of a complex, multi-instrument song.", "Solo Piano Summary": "**Action: AI Summary.** First converts the song to a solo piano arrangement, then creates a short, compressed summary.", "Add Drum Track": "**Action: Enhance.** Analyzes the rhythm of the MIDI and automatically generates a basic drum track to accompany it." } # --- Define a constant for the fallback preset name --- # This prevents errors if the preset name is changed in the dictionary. FALLBACK_PRESET_NAME = "Generic Chiptune Loop" # --- Data structure for 8-bit synthesizer presets --- # Comprehensive preset dictionary with new FX parameters for all presets # Comprehensive preset dictionary including new JRPG and Handheld classics # Note: Vibrato depth is mapped to a representative value on the 0-50 Hz slider. S8BIT_PRESETS = { # --- Classic Chiptune --- "Mario (Super Mario Bros / スーパーマリオブラザーズ)": { # Description: A bright square wave with a per-note vibrato, producing the classic bouncy platformer sound. 'waveform_type': 'Square', 'pulse_width': 0.3, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.25, 'vibrato_rate': 5.0, 'vibrato_depth': 5, 'smooth_notes_level': 0.8, 'continuous_vibrato_level': 0.25, 'bass_boost_level': 0.2, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Mega Man (Rockman / ロックマン)": { # Description: A thin, sharp square wave lead with fast vibrato, iconic for its driving, heroic melodies. 'waveform_type': 'Square', 'pulse_width': 0.2, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.15, 'vibrato_rate': 6.0, 'vibrato_depth': 8, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.85, 'bass_boost_level': 0.3, 'noise_level': 0.0, 'distortion_level': 0.05, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Zelda (The Legend of Zelda / ゼルダの伝説)": { # Description: The classic pure triangle wave lead, perfect for heroic and adventurous overworld themes. 'waveform_type': 'Triangle', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.3, 'vibrato_rate': 4.5, 'vibrato_depth': 4, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.9, 'bass_boost_level': 0.15, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Kirby's Bubbly Melody (Hoshi no Kirby / 星のカービィ)": { # Description: A soft, round square wave with a bouncy vibrato, creating a cheerful and adorable sound. 'waveform_type': 'Square', 'pulse_width': 0.4, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.2, 'vibrato_rate': 6.0, 'vibrato_depth': 4, 'smooth_notes_level': 0.85, 'continuous_vibrato_level': 0.3, # Formerly False (0.0); adds a hint of continuity for more liveliness. 'bass_boost_level': 0.1, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Pokémon (Game Boy Classics / ポケットモンスター)": { # Description: A full, friendly square wave sound, capturing the cheerful and adventurous spirit of early handheld RPGs. 'waveform_type': 'Square', 'pulse_width': 0.5, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.22, 'vibrato_rate': 5.0, 'vibrato_depth': 5, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.9, 'bass_boost_level': 0.25, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Castlevania (Akumajō Dracula / 悪魔城ドラキュラ)": { # Description: A sharp square wave with dramatic vibrato, ideal for fast, gothic, and baroque-inspired melodies. 'waveform_type': 'Square', 'pulse_width': 0.25, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.18, 'vibrato_rate': 6.5, 'vibrato_depth': 6, 'smooth_notes_level': 0.85, 'continuous_vibrato_level': 0.85, 'bass_boost_level': 0.35, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Final Fantasy (Arpeggio / ファイナルファンタジー)": { # Description: A perfect, clean square wave with zero vibrato, creating the iconic, crystal-clear arpeggio sound. 'waveform_type': 'Square', 'pulse_width': 0.5, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.22, 'vibrato_rate': 5.0, 'vibrato_depth': 0, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.2, 'bass_boost_level': 0.2, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "ONI V (Wafu Mystic / ONI V 隠忍を継ぐ者)": { # Description: A solemn triangle wave with a slow, expressive vibrato, evoking the mysterious atmosphere of Japanese folklore. 'waveform_type': 'Triangle', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.4, 'vibrato_rate': 3.5, 'vibrato_depth': 3, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.85, 'bass_boost_level': 0.4, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, # --- Advanced System Impressions --- "Commodore 64 (SID Feel)": { # Description: (Impression) Uses high-speed, shallow vibrato to mimic the characteristic "buzzy" texture of the SID chip's PWM. 'waveform_type': 'Square', 'pulse_width': 0.25, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.25, 'vibrato_rate': 8.0, 'vibrato_depth': 4, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.3, 'bass_boost_level': 0.2, 'noise_level': 0.05, 'distortion_level': 0.1, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Megadrive/Genesis (FM Grit)": { # Description: (Impression) Uses FM, distortion, and noise to capture the gritty, metallic, and aggressive tone of the YM2612 chip. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.18, 'vibrato_rate': 0.0, 'vibrato_depth': 0, 'smooth_notes_level': 0.0, 'continuous_vibrato_level': 0.9, 'bass_boost_level': 0.4, 'noise_level': 0.1, 'distortion_level': 0.2, 'fm_modulation_depth': 0.2, 'fm_modulation_rate': 150 }, "PC-98 (Touhou Feel / 東方Project)": { # Description: (Impression) A very sharp square wave with fast FM, emulating the bright, high-energy leads of Japanese PC games. 'waveform_type': 'Square', 'pulse_width': 0.15, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.12, 'vibrato_rate': 7.5, 'vibrato_depth': 7, 'smooth_notes_level': 0.95, 'continuous_vibrato_level': 0.85, 'bass_boost_level': 0.3, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.1, 'fm_modulation_rate': 200 }, "Roland SC-88 (GM Vibe)": { # Description: (Impression) A clean, stable triangle wave with no effects, mimicking the polished, sample-based sounds of General MIDI. 'waveform_type': 'Triangle', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.35, 'vibrato_rate': 0, 'vibrato_depth': 0, 'smooth_notes_level': 1.0, 'continuous_vibrato_level': 0.0, 'bass_boost_level': 0.1, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, # --- Action & Rock Leads --- "Falcom Ys (Rock Lead / イース)": { # Description: A powerful sawtooth with slight distortion, emulating the driving rock organ and guitar leads of action JRPGs. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.15, 'vibrato_rate': 5.5, 'vibrato_depth': 6, 'smooth_notes_level': 0.85, 'continuous_vibrato_level': 0.8, 'bass_boost_level': 0.4, 'noise_level': 0.05, 'distortion_level': 0.15, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Arcade Brawler Lead (Street Fighter / ストリートファイター)": { # Description: A gritty sawtooth lead with a hard attack, capturing the high-energy feel of classic fighting games. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.15, 'vibrato_rate': 5.0, 'vibrato_depth': 6, 'smooth_notes_level': 0.8, 'continuous_vibrato_level': 0.7, 'bass_boost_level': 0.4, 'noise_level': 0.05, 'distortion_level': 0.1, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Rhythm Pop Lead (Rhythm Tengoku / リズム天国)": { # Description: A clean, round square wave perfect for the snappy, catchy feel of rhythm games. 'waveform_type': 'Square', 'pulse_width': 0.5, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.18, 'vibrato_rate': 4.5, 'vibrato_depth': 4, 'smooth_notes_level': 0.9, # Formerly True -> 1.0; slightly reduced for a bit more attack. 'continuous_vibrato_level': 0.8, # Formerly True -> 1.0; slightly weakened for more defined note transitions. 'bass_boost_level': 0.3, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, # --- Epic & Orchestral Pads --- "Dragon Quest (Orchestral Feel / ドラゴンクエスト)": { # Description: A pure triangle wave with a long decay, mimicking the grand, orchestral feel of a classical flute or string section. 'waveform_type': 'Triangle', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.6, 'vibrato_rate': 3.0, 'vibrato_depth': 4, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.9, 'bass_boost_level': 0.3, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Mystic Mana Pad (Secret of Mana / 聖剣伝説2)": { # Description: A warm, ethereal square wave pad with slow vibrato, capturing a feeling of fantasy and wonder. 'waveform_type': 'Square', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.5, 'vibrato_rate': 2.5, 'vibrato_depth': 4, 'smooth_notes_level': 1.0, 'continuous_vibrato_level': 0.95, 'bass_boost_level': 0.3, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Modern JRPG Pad (Persona / ペルソナ)": { # Description: A warm, stylish square wave pad, capturing the modern, pop/jazz-infused feel of the Persona series. 'waveform_type': 'Square', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.5, 'vibrato_rate': 2.5, 'vibrato_depth': 4, 'smooth_notes_level': 1.0, 'continuous_vibrato_level': 0.95, 'bass_boost_level': 0.3, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Tactical Brass (Fire Emblem / ファイアーエムブレム)": { # Description: A powerful, sustained sawtooth emulating the bold, heroic synth-brass of Fire Emblem's tactical themes. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.4, 'vibrato_rate': 3.5, 'vibrato_depth': 5, 'smooth_notes_level': 0.95, 'continuous_vibrato_level': 0.9, 'bass_boost_level': 0.5, 'noise_level': 0.1, 'distortion_level': 0.15, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Mecha & Tactics Brass (Super Robot Wars / スーパーロボット大戦)": { # Description: A powerful, sustained sawtooth emulating the bold, heroic synth-brass of strategy and mecha anime themes. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.4, 'vibrato_rate': 3.5, 'vibrato_depth': 5, 'smooth_notes_level': 0.95, 'continuous_vibrato_level': 0.9, 'bass_boost_level': 0.5, 'noise_level': 0.1, 'distortion_level': 0.15, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Dark/Boss Atmosphere (Shin Megami Tensei / 真・女神転生)": { # Description: An aggressive sawtooth, inspired by the dark, rock-infused themes of SMT. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.35, 'vibrato_rate': 7.0, 'vibrato_depth': 12, 'smooth_notes_level': 0.1, 'continuous_vibrato_level': 0.0, 'bass_boost_level': 0.4, 'noise_level': 0.15, 'distortion_level': 0.25, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, # --- Vocal Synthesis --- "8-Bit Vocal Lead": { # Description: A soft, sustained triangle wave with gentle vibrato to mimic a singing voice. 'waveform_type': 'Triangle', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.8, 'vibrato_rate': 5.5, 'vibrato_depth': 4, # Mapped from the suggested 0.15 range 'bass_boost_level': 0.1, 'smooth_notes_level': 0.85, 'continuous_vibrato_level': 0.9, 'noise_level': 0.02, 'distortion_level': 0.0, 'fm_modulation_depth': 0.05, 'fm_modulation_rate': 20 }, "8-Bit Male Vocal": { # Description: A deeper, fuller triangle wave with more bass and slower vibrato for a masculine feel. 'waveform_type': 'Triangle', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 1.0, 'vibrato_rate': 5.0, 'vibrato_depth': 3, # Mapped from the suggested 0.12 range 'bass_boost_level': 0.3, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.85, 'noise_level': 0.015, 'distortion_level': 0.0, 'fm_modulation_depth': 0.08, 'fm_modulation_rate': 25 }, "8-Bit Female Vocal": { # Description: A brighter, lighter triangle wave with faster vibrato and less bass for a feminine feel. 'waveform_type': 'Triangle', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.7, 'vibrato_rate': 6.0, 'vibrato_depth': 5, # Mapped from the suggested 0.18 range 'bass_boost_level': 0.05, 'smooth_notes_level': 0.85, 'continuous_vibrato_level': 0.92, 'noise_level': 0.025, 'distortion_level': 0.0, 'fm_modulation_depth': 0.04, 'fm_modulation_rate': 30 }, "Lo-Fi Vocal": { # Description: A gritty, noisy square wave with a short decay to simulate a low-resolution vocal sample. 'waveform_type': 'Square', 'pulse_width': 0.48, 'envelope_type': 'Plucky (AD Envelope)', # "Short" implies a plucky, not sustained, envelope 'decay_time_s': 0.4, 'vibrato_rate': 4.8, 'vibrato_depth': 2, # Mapped from the suggested 0.10 range 'bass_boost_level': 0.1, 'smooth_notes_level': 0.65, 'continuous_vibrato_level': 0.6, 'noise_level': 0.05, 'distortion_level': 0.05, 'fm_modulation_depth': 0.02, 'fm_modulation_rate': 20 }, # --- Sound FX & Experimental --- "Sci-Fi Energy Field": { # Description: (SFX) High-speed vibrato and noise create a constant, shimmering hum suitable for energy shields or force fields. 'waveform_type': 'Triangle', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.4, 'vibrato_rate': 10.0, 'vibrato_depth': 3, 'smooth_notes_level': 0.85, 'continuous_vibrato_level': 0.9, 'bass_boost_level': 0.1, 'noise_level': 0.1, 'distortion_level': 0.0, 'fm_modulation_depth': 0.05, 'fm_modulation_rate': 50 }, "Industrial Alarm": { # Description: (SFX) Extreme vibrato rate on a sawtooth wave produces a harsh, metallic, dissonant alarm sound. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.2, 'vibrato_rate': 15.0, 'vibrato_depth': 8, 'smooth_notes_level': 0.0, 'continuous_vibrato_level': 0.0, 'bass_boost_level': 0.3, 'noise_level': 0.2, 'distortion_level': 0.3, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Laser Charge-Up": { # Description: (SFX) Extreme vibrato depth creates a dramatic, rising pitch effect, perfect for sci-fi weapon sounds. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.3, 'vibrato_rate': 4.0, 'vibrato_depth': 25, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.95, 'bass_boost_level': 0.2, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, "Unstable Machine Core": { # Description: (SFX) Maximum depth and distortion create a chaotic, atonal noise, simulating a machine on the verge of exploding. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Sustained (Full Decay)', 'decay_time_s': 0.5, 'vibrato_rate': 1.0, 'vibrato_depth': 50, 'smooth_notes_level': 0.0, 'continuous_vibrato_level': 0.9, 'bass_boost_level': 0.5, 'noise_level': 0.3, 'distortion_level': 0.4, 'fm_modulation_depth': 0.5, 'fm_modulation_rate': 10 }, "Hardcore Gabber Kick": { # Description: (Experimental) Maximum bass boost and distortion create an overwhelmingly powerful, clipped kick drum sound. 'waveform_type': 'Sawtooth', 'pulse_width': 0.5, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.1, 'vibrato_rate': 0, 'vibrato_depth': 0, 'smooth_notes_level': 0.0, 'continuous_vibrato_level': 0.0, 'bass_boost_level': 0.8, 'noise_level': 0.2, 'distortion_level': 0.5, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, # --- Utility & Starting Points --- "Generic Chiptune Loop": { # Description: A well-balanced, pleasant square wave lead that serves as a great starting point for custom sounds. 'waveform_type': 'Square', 'pulse_width': 0.25, 'envelope_type': 'Plucky (AD Envelope)', 'decay_time_s': 0.2, 'vibrato_rate': 5.5, 'vibrato_depth': 4, 'smooth_notes_level': 0.9, 'continuous_vibrato_level': 0.85, 'bass_boost_level': 0.25, 'noise_level': 0.0, 'distortion_level': 0.0, 'fm_modulation_depth': 0.0, 'fm_modulation_rate': 0.0 }, } # --- Data structure for basic_pitch transcription presets --- BASIC_PITCH_PRESETS = { # --- General & All-Purpose --- "Default (Balanced)": { 'description': "A good all-around starting point for most music types.", 'onset_threshold': 0.5, 'frame_threshold': 0.3, 'minimum_note_length': 128, 'minimum_frequency': 60, 'maximum_frequency': 4000, 'infer_onsets': True, 'melodia_trick': True, 'multiple_bends': False }, "Anime / J-Pop": { 'description': "For tracks with clear melodies and pop/rock arrangements.", 'onset_threshold': 0.5, 'frame_threshold': 0.3, 'minimum_note_length': 150, 'minimum_frequency': 40, 'maximum_frequency': 2500, 'infer_onsets': True, 'melodia_trick': True, 'multiple_bends': True }, # --- Specific Instruments --- "Solo Vocals": { 'description': "Optimized for a single singing voice. Sensitive to nuances.", 'onset_threshold': 0.4, 'frame_threshold': 0.3, 'minimum_note_length': 100, 'minimum_frequency': 80, 'maximum_frequency': 1200, 'infer_onsets': True, 'melodia_trick': True, 'multiple_bends': True }, "Solo Piano": { 'description': "For solo piano with a wide dynamic and frequency range.", 'onset_threshold': 0.4, 'frame_threshold': 0.3, 'minimum_note_length': 120, 'minimum_frequency': 27, 'maximum_frequency': 4200, 'infer_onsets': True, 'melodia_trick': True, 'multiple_bends': True }, "Acoustic Guitar": { 'description': "Balanced for picked or strummed acoustic guitar.", 'onset_threshold': 0.5, 'frame_threshold': 0.3, 'minimum_note_length': 90, 'minimum_frequency': 80, 'maximum_frequency': 2500, 'infer_onsets': True, 'melodia_trick': True, 'multiple_bends': False }, "Bass Guitar": { 'description': "Isolates and transcribes only the low frequencies of a bassline.", 'onset_threshold': 0.4, 'frame_threshold': 0.3, 'minimum_note_length': 100, 'minimum_frequency': 30, 'maximum_frequency': 400, 'infer_onsets': True, 'melodia_trick': True, 'multiple_bends': False }, "Percussion / Drums": { 'description': "For drums and rhythmic elements. Catches fast, sharp hits.", 'onset_threshold': 0.7, 'frame_threshold': 0.6, 'minimum_note_length': 30, 'minimum_frequency': 40, 'maximum_frequency': 10000, 'infer_onsets': True, 'melodia_trick': False, 'multiple_bends': False }, # --- Complex Genres --- "Rock / Metal": { 'description': "Higher thresholds for distorted guitars, bass, and drums in a dense mix.", 'onset_threshold': 0.6, 'frame_threshold': 0.4, 'minimum_note_length': 100, 'minimum_frequency': 50, 'maximum_frequency': 3000, 'infer_onsets': True, 'melodia_trick': True, 'multiple_bends': True }, "Jazz (Multi-instrument)": { 'description': "High thresholds to separate notes in complex, improvisational passages.", 'onset_threshold': 0.7, 'frame_threshold': 0.5, 'minimum_note_length': 150, 'minimum_frequency': 55, 'maximum_frequency': 2000, 'infer_onsets': True, 'melodia_trick': False, 'multiple_bends': True }, "Classical (Orchestral)": { 'description': "Longer note length to focus on sustained notes and filter out performance noise.", 'onset_threshold': 0.5, 'frame_threshold': 0.4, 'minimum_note_length': 200, 'minimum_frequency': 32, 'maximum_frequency': 4200, 'infer_onsets': True, 'melodia_trick': True, 'multiple_bends': True }, "Electronic / Synth": { 'description': "Low thresholds and short note length for sharp, synthetic sounds.", 'onset_threshold': 0.3, 'frame_threshold': 0.2, 'minimum_note_length': 50, 'minimum_frequency': 20, 'maximum_frequency': 8000, 'infer_onsets': True, 'melodia_trick': False, 'multiple_bends': False } } # --- UI visibility logic now controls three components --- def update_vocal_ui_visibility(separate_vocals): """Shows or hides the separation-related UI controls based on selections.""" is_visible = gr.update(visible=separate_vocals) return is_visible, is_visible, is_visible def update_ui_visibility(transcription_method, soundfont_choice): """ Dynamically updates the visibility of UI components based on user selections. """ is_general = (transcription_method == "General Purpose") is_8bit = (soundfont_choice == SYNTH_8_BIT_LABEL) return { general_transcription_settings: gr.update(visible=is_general), synth_8bit_settings: gr.update(visible=is_8bit), } # --- Function to control visibility of advanced MIDI rendering options --- def update_advanced_midi_options_visibility(render_type_choice): """ Shows or hides the advanced MIDI rendering options based on the render type. The options are only visible if the type is NOT 'Render as-is'. """ is_visible = (render_type_choice != "Render as-is") return gr.update(visible=is_visible) # --- UI controller function to update the description text --- def update_render_type_description(render_type_choice): """ Returns the description for the selected render type. """ return RENDER_TYPE_DESCRIPTIONS.get(render_type_choice, "Select a render type to see its description.") # --- Controller function to apply basic_pitch presets to the UI --- def apply_basic_pitch_preset(preset_name): if preset_name not in BASIC_PITCH_PRESETS: # If "Custom" is selected or name is invalid, don't change anything return {comp: gr.update() for comp in basic_pitch_ui_components} settings = BASIC_PITCH_PRESETS[preset_name] # Return a dictionary that maps each UI component to its new value return { onset_threshold: gr.update(value=settings['onset_threshold']), frame_threshold: gr.update(value=settings['frame_threshold']), minimum_note_length: gr.update(value=settings['minimum_note_length']), minimum_frequency: gr.update(value=settings['minimum_frequency']), maximum_frequency: gr.update(value=settings['maximum_frequency']), infer_onsets: gr.update(value=settings['infer_onsets']), melodia_trick: gr.update(value=settings['melodia_trick']), multiple_pitch_bends: gr.update(value=settings['multiple_bends']) } # --- Function to apply 8-bit synthesizer presets --- # --- This function must be defined before the UI components that use it --- def apply_8bit_preset(preset_name): """ Takes the name of a preset and returns a dictionary of gr.update objects to set the values of the 13 8-bit synthesizer control components. This version is more robust as it directly maps keys to UI components. """ # If a special value is selected or the preset is not found, return empty updates for all controls. if preset_name in ["Custom", "Auto-Recommend (Analyze MIDI)"] or preset_name not in S8BIT_PRESETS: # We create a dictionary mapping each control component to an empty update. s8bit_control_keys = [key for key in ALL_PARAM_KEYS if key.startswith('s8bit_') and key != 's8bit_preset_selector'] return {ui_component_map[key]: gr.update() for key in s8bit_control_keys} # Get the settings dictionary for the chosen preset. settings = S8BIT_PRESETS[preset_name] updates = {} # Iterate through the KEY-VALUE pairs in the chosen preset's settings. for simple_key, value in settings.items(): # Reconstruct the full component key (e.g., 'waveform_type' -> 's8bit_waveform_type') full_key = f"s8bit_{simple_key}" # Check if this key corresponds to a valid UI component if full_key in ui_component_map: component = ui_component_map[full_key] updates[component] = gr.update(value=value) return updates # --- UI Controller Function for Dynamic Visibility --- def update_separation_mode_ui(is_advanced): """ Updates the visibility and labels of UI components based on whether the advanced separation mode is enabled. """ if is_advanced: # Advanced Mode: Show individual controls, label becomes "Other" return { advanced_separation_controls: gr.update(visible=True), transcribe_drums: gr.update(visible=True), transcribe_bass: gr.update(visible=True), transcribe_other_or_accompaniment: gr.update(label="Transcribe Other"), merge_drums_to_render: gr.update(visible=True), merge_bass_to_render: gr.update(visible=True), merge_other_or_accompaniment: gr.update(label="Merge Other") } else: # Simple Mode: Hide individual controls, label becomes "Accompaniment" return { advanced_separation_controls: gr.update(visible=False), transcribe_drums: gr.update(visible=False), transcribe_bass: gr.update(visible=False), transcribe_other_or_accompaniment: gr.update(label="Transcribe Accompaniment"), merge_drums_to_render: gr.update(visible=False), merge_bass_to_render: gr.update(visible=False), merge_other_or_accompaniment: gr.update(label="Merge Accompaniment") } # --- Use the dataclass to define the master list of parameter keys --- # This is now the single source of truth for parameter order. ALL_PARAM_KEYS = [field.name for field in fields(AppParameters) if field.name not in ["input_file", "batch_input_files"]] app = gr.Blocks(theme=gr.themes.Base()) with app: gr.Markdown("