# ================================================================= # # 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 remerge_vocals: bool = False transcription_target: str = "Transcribe Music (Accompaniment)" transcribe_both_stems: 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 # ================================================================================================= # === Helper Functions === # ================================================================================================= 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): """ 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. """ 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 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 osc_phase[i] = 0.0 # Independent phase tracking for each instrument for note in instrument.notes: 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 (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 note_waveform = signal.sawtooth(phase, width=0.5) # --- 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 Simulation (White Noise) --- if params.s8bit_noise_level > 0: note_waveform += np.random.uniform(-1, 1, num_samples) * params.s8bit_noise_level # --- Distortion (Wave Shaping) --- if params.s8bit_distortion_level > 0: # Using a tanh function for a smoother, "warmer" distortion note_waveform = np.tanh(note_waveform * (1 + params.s8bit_distortion_level * 5)) # --- ADSR Envelope --- start_amp = note.velocity / 127.0 envelope = np.zeros(num_samples) if params.s8bit_envelope_type == 'Plucky (AD Envelope)': attack_samples = min(int(0.005 * fs), num_samples) 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) # --- Graded Note Smoothing --- # The level controls the length of the fade in/out. Max fade is 10ms. if params.s8bit_smooth_notes_level > 0 and num_samples > 10: fade_length = int(fs * 0.01 * params.s8bit_smooth_notes_level) fade_samples = min(fade_length, num_samples // 2) if fade_samples > 0: envelope[:fade_samples] *= np.linspace(0.5, 1.0, fade_samples) envelope[-fade_samples:] *= np.linspace(1.0, 0.0, fade_samples) # Apply envelope to the (potentially combined) waveform note_waveform *= envelope 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 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 # ================================================================================================= # === 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): """ 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) # 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}" # This will store the other part if separation is performed other_part_tensor = None other_part_sr = None # --- 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.") 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, now decides which stem to process --- if not params.separate_vocals or demucs_model is None: if params.separate_vocals and demucs_model is None: print("ERROR: Demucs model not loaded. Skipping 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) else: # --- Vocal Separation Workflow --- update_progress(0.2, "Separating vocals with Demucs...") # Convert to a common format (stereo, float32) that demucs expects audio_tensor = convert_audio(audio_tensor, native_sample_rate, demucs_model.samplerate, demucs_model.audio_channels) 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.") # --- Robust stem handling to prevent CUDA errors --- # Instead of complex GPU indexing, we create a dictionary of stems on the CPU. # This is safer and more robust across different hardware. sources = {} for i, source_name in enumerate(demucs_model.sources): sources[source_name] = all_stems[i] vocals_tensor = sources['vocals'] # Sum the other stems to create the accompaniment. # This loop is safer than a single complex indexing operation. accompaniment_tensor = torch.zeros_like(vocals_tensor) for source_name, stem_tensor in sources.items(): if source_name != 'vocals': accompaniment_tensor += stem_tensor # --- Save both stems to temporary files --- vocals_path = os.path.join(temp_dir, f"{base_name}_vocals.flac") accompaniment_path = os.path.join(temp_dir, f"{base_name}_accompaniment.flac") torchaudio.save(vocals_path, vocals_tensor.cpu(), demucs_model.samplerate) torchaudio.save(accompaniment_path, accompaniment_tensor.cpu(), demucs_model.samplerate) # --- Determine which stem is the primary target and which is the "other part" --- primary_target_path = vocals_path if params.transcription_target == "Transcribe Vocals" else accompaniment_path other_part_path = accompaniment_path if params.transcription_target == "Transcribe Vocals" else vocals_path # Store the audio tensor of the "other part" for potential audio re-merging other_part_tensor = accompaniment_tensor if params.transcription_target == "Transcribe Vocals" else vocals_tensor other_part_sr = demucs_model.samplerate print("Separation complete.") # --- Main Branching Logic: Transcribe one or both stems --- if not params.transcribe_both_stems: print(f"Transcribing primary target only: {os.path.basename(primary_target_path)}") update_progress(0.4, f"Transcribing primary target: {os.path.basename(primary_target_path)}") midi_path_for_rendering = _transcribe_stem(primary_target_path, os.path.splitext(os.path.basename(primary_target_path))[0], temp_dir, params) else: print("Transcribing BOTH stems and merging the MIDI results.") # Transcribe the primary target update_progress(0.4, "Transcribing primary stem...") midi_path_primary = _transcribe_stem(primary_target_path, os.path.splitext(os.path.basename(primary_target_path))[0], temp_dir, params) # Transcribe the other part update_progress(0.5, "Transcribing second stem...") midi_path_other = _transcribe_stem(other_part_path, os.path.splitext(os.path.basename(other_part_path))[0], temp_dir, params) # Merge the two resulting MIDI files if midi_path_primary and midi_path_other: update_progress(0.55, "Merging transcribed MIDIs...") final_merged_midi_path = os.path.join(temp_dir, f"{base_name}_full_transcription.mid") print(f"Merging transcribed MIDI files into {os.path.basename(final_merged_midi_path)}") # A more robust MIDI merge is needed here primary_midi = pretty_midi.PrettyMIDI(midi_path_primary) other_midi = pretty_midi.PrettyMIDI(midi_path_other) # Add all instruments from the other midi to the primary one for instrument in other_midi.instruments: instrument.name = f"Other - {instrument.name}" # Rename to avoid confusion primary_midi.instruments.append(instrument) primary_midi.write(final_merged_midi_path) midi_path_for_rendering = final_merged_midi_path elif midi_path_primary: print("Warning: Transcription of the 'other' part failed. Using primary transcription only.") midi_path_for_rendering = midi_path_primary else: raise gr.Error("Transcription of the primary target failed. Aborting.") 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}.") 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) # --- Vocal Re-merging Logic --- # Vocal Re-merging only happens for audio files, so its progress value doesn't need to be conditional. if params.separate_vocals and params.remerge_vocals and not params.transcribe_both_stems and other_part_tensor is not None: update_progress(0.8, "Re-merging rendered audio with vocals...") print(f"Re-merging the non-transcribed part with newly rendered music...") # 1. Unpack the original rendered audio from the results rendered_srate, rendered_music_int16 = results_tuple[4] # 2. Convert the rendered music to a float tensor rendered_music_float = rendered_music_int16.astype(np.float32) / 32767.0 rendered_music_tensor = torch.from_numpy(rendered_music_float).T # 3. Resample if necessary if rendered_srate != other_part_sr: resampler = torchaudio.transforms.Resample(rendered_srate, other_part_sr) rendered_music_tensor = resampler(rendered_music_tensor) # 4. Pad to match lengths len_music = rendered_music_tensor.shape[1] len_other = other_part_tensor.shape[1] if len_music > len_other: padding = len_music - len_other other_part_tensor = torch.nn.functional.pad(other_part_tensor, (0, padding)) elif len_other > len_music: padding = len_other - len_music rendered_music_tensor = torch.nn.functional.pad(rendered_music_tensor, (0, padding)) # 5. Merge and normalize merged_audio_tensor = rendered_music_tensor + other_part_tensor.cpu() max_abs = torch.max(torch.abs(merged_audio_tensor)) if max_abs > 1.0: merged_audio_tensor /= max_abs # 6. Convert back to the required format (int16 numpy array) merged_audio_int16 = (merged_audio_tensor.T.numpy() * 32767).astype(np.int16) # 7. Create the new audio tuple and UPDATE the main results_tuple new_audio_tuple = (other_part_sr, merged_audio_int16) temp_results_list = list(results_tuple) temp_results_list[4] = new_audio_tuple results_tuple = tuple(temp_results_list) # results_tuple is now updated print("Re-merging complete.") # --- Save final audio and return path --- update_progress(0.9, "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 === # ================================================================================================= # --- 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) # --- Nested Progress Logic --- # Define a local function to scale the sub-progress of the pipeline # into the correct slot of the main batch progress bar. def batch_progress_updater(local_fraction, desc): # Calculate the overall progress based on which file we are on (i) # and the progress within that file (local_fraction). progress_per_file = 1 / total_files overall_fraction = (i / total_files) + (local_fraction * progress_per_file) progress(overall_fraction, desc=f"({i+1}/{total_files}) {filename}: {desc}") progress(i / total_files, desc=f"Processing {os.path.basename(input_path)} ({i+1}/{total_files})") # --- Pass the batch_timestamp to the pipeline --- results, _ = run_single_file_pipeline(input_path, batch_timestamp, params, progress=batch_progress_updater) 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: final_ui_updates.append(getattr(final_params, key)) # 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 # --- 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("

Audio-to-MIDI & Advanced Renderer

") gr.Markdown( "**Upload a Audio for transcription-then-rendering, or a MIDI for rendering-only.**\n\n" "This application combines piano audio transcription with a powerful MIDI transformation and rendering toolkit. " "Based on the work of [asigalov61](https://github.com/asigalov61)." ) # --- Use Tabs for different workflows --- with gr.Tabs(): waveform_options = gr.WaveformOptions(show_recording_waveform=False) # --- TAB 1: SINGLE FILE PROCESSING --- with gr.TabItem("Single File Processing"): # --- All of your existing UI components go inside this Tab --- with gr.Row(): with gr.Column(scale=1): # --- INPUT COLUMN --- gr.Markdown("## 1. Upload File") # Changed from gr.File to gr.Audio to allow for audio preview. # type="filepath" ensures the component returns a string path to the uploaded file. # The component will show a player for supported audio types (e.g., WAV, MP3). input_file = gr.Audio( label="Input Audio or MIDI File", type="filepath", sources=["upload"], waveform_options=waveform_options ) # --- The single file processing button --- submit_btn = gr.Button("Process and Render Single File", variant="primary") with gr.Column(scale=2): # --- OUTPUT COLUMN --- gr.Markdown("### 2. Results") output_midi_title = gr.Textbox(label="MIDI Title") output_song_description = gr.Textbox(label="MIDI Description", lines=3) output_audio = gr.Audio(label="Rendered Audio Output", format="wav", waveform_options=waveform_options) output_plot = gr.Plot(label="MIDI Score Plot") with gr.Row(): output_midi = gr.File(label="Download Processed MIDI File", file_types=[".mid"]) output_midi_md5 = gr.Textbox(label="Output MIDI MD5 Hash") output_midi_summary = gr.Textbox(label="MIDI metadata summary", lines=4) # --- TAB 2: BATCH PROCESSING --- with gr.TabItem("Batch Processing"): with gr.Row(): with gr.Column(): gr.Markdown("### 1. Upload Files") gr.Markdown("Uses the **global settings** configured above.") batch_input_files = gr.File( label="Upload Audio or MIDI Files", file_count="multiple" ) batch_process_btn = gr.Button("Process Batch", variant="primary") with gr.Column(): gr.Markdown("### 2. Download Results") batch_output_audio_files = gr.File( label="Download Rendered FLAC Files", file_count="multiple", interactive=False ) batch_output_midi_files = gr.File( label="Download Processed MIDI Files", file_count="multiple", interactive=False ) # --- Global Settings Accordion, Define all settings in a global, shared accordion --- with gr.Accordion("▶️ Configure Global Settings (for both Single File and Batch)", open=True): with gr.Row(): with gr.Column(scale=1): gr.Markdown("### Transcription Settings") # --- Transcription Method Selector --- transcription_method = gr.Radio(["General Purpose", "Piano-Specific"], label="Audio Transcription Method", value="General Purpose", info="Choose 'General Purpose' for most music (vocals, etc.). Choose 'Piano-Specific' only for solo piano recordings.") # --- Stereo Processing Checkbox --- enable_stereo_processing = gr.Checkbox(label="Enable Stereo Transcription", value=False, info="If checked, left/right audio channels are transcribed separately and merged. Doubles processing time.") # --- Vocal Separation Checkboxes --- with gr.Group(): separate_vocals = gr.Checkbox(label="Separate Vocals", value=False, info="If checked, separates the audio into vocals and music stems before processing.") transcription_target = gr.Radio(["Transcribe Music (Accompaniment)", "Transcribe Vocals"], label="Transcription Target", value="Transcribe Music (Accompaniment)", visible=False, info="Choose which part of the separated audio to transcribe to MIDI.") remerge_vocals = gr.Checkbox(label="Re-merge Other Part with Rendered Audio", value=False, visible=False, info="After rendering, merges the non-transcribed part (e.g., original vocals) back with the new music.") transcribe_both_stems = gr.Checkbox(label="Transcribe Both Parts & Merge MIDI", value=False, visible=False, info="If checked, transcribes BOTH vocals and music, then merges them into one MIDI file for rendering. Disables audio re-merging.") with gr.Accordion("General Purpose Transcription Settings", open=True) as general_transcription_settings: # --- Preset dropdown for basic_pitch --- basic_pitch_preset_selector = gr.Dropdown( choices=["Custom"] + list(BASIC_PITCH_PRESETS.keys()), value="Default (Balanced)", label="Transcription Profile Preset", info="Select a profile to auto-fill settings for different instrument types." "For reference only; it is recommended to test and adjust for optimal results.") # --- The existing basic_pitch components --- onset_threshold = gr.Slider(0.0, 1.0, value=0.5, step=0.05, label="On-set Threshold", info="Sensitivity for detecting note beginnings. Higher is stricter.") frame_threshold = gr.Slider(0.0, 1.0, value=0.3, step=0.05, label="Frame Threshold", info="Sensitivity for detecting active notes. Higher is stricter.") minimum_note_length = gr.Slider(10, 500, value=128, step=1, label="Minimum Note Length (ms)", info="Filters out very short, noisy notes.") minimum_frequency = gr.Slider(0, 500, value=60, step=5, label="Minimum Frequency (Hz)", info="Ignores pitches below this frequency.") maximum_frequency = gr.Slider(501, 10000, value=4000, step=10, label="Maximum Frequency (Hz)", info="Ignores pitches above this frequency.") infer_onsets = gr.Checkbox(value=True, label="Infer Onsets (Boost Onsets)") melodia_trick = gr.Checkbox(value=True, label="Melodia Trick (Contour Optimization)") multiple_pitch_bends = gr.Checkbox(value=False, label="Allow Multiple Pitch Bends") with gr.Column(scale=1): # --- Rendering Settings --- gr.Markdown("### MIDI Transformation & Rendering Settings") render_type = gr.Radio( list(RENDER_TYPE_DESCRIPTIONS.keys()), # Use keys from dict for choices ["Render as-is", "Custom render", "Extract melody", "Flip", "Reverse", "Repair Durations", "Repair Chords", "Remove Duplicate Pitches", "Longest Repeating Phrase", "Multi-Instrumental Summary", "Solo Piano Summary", "Add Drum Track"], label="MIDI Transformation Render Type", value="Render as-is", info="Apply transformations to the MIDI before rendering. Select 'Render as-is' for basic rendering or other options for transformations.") # --- A Markdown box for the dynamic descriptions --- render_type_info = gr.Markdown( value=RENDER_TYPE_DESCRIPTIONS["Render as-is"], # Set initial value elem_classes="description-box" # Optional: for CSS styling ) # --- SoundFont Bank with 8-bit option --- soundfont_bank = gr.Dropdown( [SYNTH_8_BIT_LABEL] + list(soundfonts_dict.keys()), label="SoundFont / Synthesizer", value=list(soundfonts_dict.keys())[0] if soundfonts_dict else SYNTH_8_BIT_LABEL) render_sample_rate = gr.Radio( ["16000", "32000", "44100"], label="Audio Sample Rate", value="44100") with gr.Accordion("Advanced MIDI Rendering Options", open=False) as advanced_rendering_options: render_with_sustains = gr.Checkbox(label="Apply sustain pedal effects (if present)", value=True, info="Applies sustain pedal effects (CC64) to lengthen notes, creating a more realistic and connected performance, especially for piano.") render_output_as_solo_piano = gr.Checkbox(label="Convert to Solo Piano (Grand Piano patch)", value=False, info="Converts all non-drum instruments to a Grand Piano patch, creating a solo piano arrangement of the entire score.") render_remove_drums = gr.Checkbox(label="Remove drum track", value=False, info="Removes the entire drum track (typically MIDI Channel 9) from the score. Ideal for creating instrumental or karaoke versions.") render_transpose_to_C4 = gr.Checkbox(label="Transpose entire score to center around C4", value=False, info="Transposes the entire score so that its average pitch is centered around C4 (MIDI note 60). Useful for standardizing key.") render_transpose_value = gr.Slider(-12, 12, value=0, step=1, label="Transpose (semitones)", info="Shifts the pitch of all non-drum notes up (positive values) or down (negative values) by the specified number of semitones.") custom_render_patch = gr.Slider(-1, 127, value=-1, step=1, label="Force MIDI Patch (-1 to disable)", info="Forces all non-drum instruments to use a single specified MIDI patch number. Set to -1 to use the original instruments.") merge_misaligned_notes = gr.Slider(-1, 127, value=-1, label="Time to merge notes in ms (-1 to disable)", info="Aligns the start times of notes that are played almost simultaneously (within the specified ms threshold). Cleans up sloppy timing. -1 to disable.") render_align = gr.Radio( ["Do not align", "Start Times", "Start Times and Durations", "Start Times and Split Durations"], label="Align notes to musical bars", value="Do not align", info="Quantizes the score to a fixed bar length. 'Start Times' aligns onsets. " "'Durations' trims notes at the bar line. 'Split Durations' splits notes that cross the bar line." ) with gr.Column(scale=1): # --- 8-bit Synthesizer Settings --- # # ================================================================================= # === 8-Bit Synthesizer Parameter Guide === # ================================================================================= # # --- Basic Tone Shaping --- # # Waveform Type: The fundamental timbre of the sound. # - Square: The classic, bright, somewhat hollow sound of the NES. Its tone is heavily modified by Pulse Width. # - Sawtooth: Aggressive, buzzy, and rich. Great for intense leads or gritty basslines. # - Triangle: Soft, pure, and flute-like. Often used for basslines or gentler melodies. # # Pulse Width (Square Wave Only): Modifies the character of the Square wave. # - Low (near 0.1) or High (near 0.9): Creates a thin, sharp, or nasal sound. A common choice for classic leads. # - Mid (near 0.5): A "perfect" square wave. The sound is full, round, and most robust. # # Envelope Type: Shapes the volume of each note over its duration. # - Plucky (AD): Creates a percussive, short sound that attacks instantly and then fades. Ideal for fast melodies and arpeggios. # - Sustained (Full Decay): Creates a held-out sound that lasts for the note's full duration. Ideal for pads and atmospheric sounds. # # Decay Time (s): Controls how long a note's sound lasts (in the Plucky envelope). # - Low: Very short, staccato notes. # - High: Longer, more resonant notes that can bleed into each other. # # Bass Boost Level: Mixes in a sub-octave (a square wave one octave lower). # - Low (or 0): The pure, original waveform. # - High: Adds significant weight, thickness, and power to the sound. # # --- Modulation & Performance --- # # Vibrato Rate (Hz): The SPEED of the pitch wobble. # - Low: A slow, gentle wavering effect. # - High (8Hz+): A fast, frantic buzzing or trembling effect. Can create "ring-mod" style sounds at extreme values. # # Vibrato Depth (Hz): The INTENSITY of the pitch wobble. # - Low (or 0): A very subtle effect, or no vibrato at all. # - High: An extreme, dramatic pitch bend. Can sound chaotic or like a siren at extreme values. # # Smooth Notes (Checkbox): # - Enabled: Applies a tiny fade-in/out to reduce clicking artifacts. Makes the sound slightly softer but cleaner. # - Disabled: More abrupt, harsh note onsets. Can be desirable for an aggressive sound. # # Continuous Vibrato (Checkbox): # - Enabled: The vibrato is smooth and connected across a musical phrase, creating a "singing" or legato effect. # - Disabled: The vibrato resets on each new note, creating a bouncy, per-note, staccato effect (key for the "Mario" style). # # --- FX & Advanced Synthesis --- # # Noise Level: Mixes in white noise with the main waveform. # - Low (or 0): No noise. # - High: Adds "air," "grit," or a "hissing" quality. Essential for simulating percussion or creating wind-like sound effects. # # Distortion Level: Applies a wave-shaping algorithm to make the sound harsher. # - Low (or 0): The clean, original sound. # - High: Progressively crushes and saturates the waveform, creating a very aggressive, "fuzzy" or "broken" tone. # # FM Depth (Frequency Modulation): Controls the intensity of the frequency modulation. # - Low (or 0): No FM effect. # - High: The main frequency is more heavily altered by the FM Rate, creating complex, bell-like, metallic, or dissonant tones. # # FM Rate (Frequency Modulation): Controls the speed of the modulating oscillator. # - Low: Creates a slow, vibrato-like or "wobbling" FM effect. # - High: Creates fast modulation, resulting in bright, complex, often metallic harmonics and sidebands. # ================================================================================= # # --- New option for auto-recommendation --- # Define the 8-bit UI components in one place for easy reference gr.Markdown("### 8-bit Synthesizer Settings") with gr.Accordion("8-bit Synthesizer Settings", open=True, visible=False) as synth_8bit_settings: s8bit_preset_selector = gr.Dropdown(choices=["Custom", "Auto-Recommend (Analyze MIDI)"] + list(S8BIT_PRESETS.keys()), value="Custom", label="Style Preset", info="Select a preset to auto-fill the settings below. Choose 'Custom' for manual control.\nFor reference and entertainment only. These presets are not guaranteed to be perfectly accurate.") s8bit_waveform_type = gr.Dropdown(['Square', 'Sawtooth', 'Triangle'], value='Square', label="Waveform Type") s8bit_pulse_width = gr.Slider(0.01, 0.99, value=0.5, step=0.01, label="Pulse Width (Square Wave Only)") s8bit_envelope_type = gr.Dropdown(['Plucky (AD Envelope)', 'Sustained (Full Decay)'], value='Plucky (AD Envelope)', label="Envelope Type") s8bit_decay_time_s = gr.Slider(0.01, 1.0, value=0.1, step=0.01, label="Decay Time (s)") s8bit_vibrato_rate = gr.Slider(0, 20, value=5, label="Vibrato Rate (Hz)") s8bit_vibrato_depth = gr.Slider(0, 50, value=0, label="Vibrato Depth (Hz)") s8bit_bass_boost_level = gr.Slider(0.0, 1.0, value=0.0, step=0.05, label="Bass Boost Level", info="Adjusts the volume of the sub-octave. 0 is off.") s8bit_smooth_notes_level = gr.Slider(0.0, 1.0, value=0.0, step=0.05, label="Smooth Notes Level", info="Level of fade-in/out to reduce clicks. 0=off, 1=max.") s8bit_continuous_vibrato_level = gr.Slider(0.0, 1.0, value=0.0, step=0.05, label="Continuous Vibrato Level", info="Controls vibrato continuity. 0=resets per note, 1=fully continuous.") # --- New accordion for advanced effects --- with gr.Accordion("Advanced Synthesis & FX", open=False): s8bit_noise_level = gr.Slider(0.0, 1.0, value=0.0, step=0.05, label="Noise Level", info="Mixes in white noise. Great for percussion or adding 'air'.") s8bit_distortion_level = gr.Slider(0.0, 0.9, value=0.0, step=0.05, label="Distortion Level", info="Applies wave-shaping distortion for a grittier, harsher sound.") s8bit_fm_modulation_depth = gr.Slider(0.0, 1.0, value=0.0, step=0.05, label="FM Depth", info="Depth of Frequency Modulation. Creates complex, metallic, or bell-like tones.") s8bit_fm_modulation_rate = gr.Slider(0.0, 500.0, value=0.0, step=1.0, label="FM Rate", info="Rate of Frequency Modulation. Higher values create brighter, more complex harmonics.") # Create a dictionary mapping key names to the actual Gradio components ui_component_map = locals() # Build the list of all setting components in the correct order using ALL_PARAM_KEYS all_settings_components = [ui_component_map[key] for key in ALL_PARAM_KEYS] # --- FIX START: Isolate the preset selector from the controls it updates --- # Original list of all 14 synth components s8bit_ui_keys = [key for key in ALL_PARAM_KEYS if key.startswith('s8bit_')] s8bit_ui_components = [ui_component_map[key] for key in s8bit_ui_keys] # NEW: Create a separate list containing only the 13 controls to be updated s8bit_control_components = [comp for comp in s8bit_ui_components if comp != s8bit_preset_selector] # The list of basic_pitch UI components that can be updated by its preset selector. basic_pitch_keys = ['onset_threshold', 'frame_threshold', 'minimum_note_length', 'minimum_frequency', 'maximum_frequency', 'infer_onsets', 'melodia_trick', 'multiple_pitch_bends'] basic_pitch_ui_components = [ui_component_map[key] for key in basic_pitch_keys] # Define inputs and outputs for Gradio events single_file_inputs = [input_file] + all_settings_components result_outputs = [output_midi_md5, output_midi_title, output_midi_summary, output_midi, output_audio, output_plot, output_song_description] # The output list for the single file process now correctly includes all 14 synth components single_file_outputs = result_outputs + s8bit_ui_components batch_inputs = [batch_input_files] + all_settings_components batch_outputs = [batch_output_audio_files, batch_output_midi_files] # Event Handling for Single File Tab submit_btn.click( fn=process_and_render_file, inputs=single_file_inputs, outputs=single_file_outputs ) # --- Event Handling for Batch Tab --- batch_process_btn.click( fn=batch_process_files, inputs=batch_inputs, outputs=batch_outputs ) # Event listeners for UI visibility and presets separate_vocals.change( fn=update_vocal_ui_visibility, inputs=separate_vocals, outputs=[transcription_target, remerge_vocals, transcribe_both_stems] ) # --- Listeners for dynamic UI updates --- transcription_method.change( fn=lambda x: gr.update(visible=(x == "General Purpose")), inputs=transcription_method, outputs=general_transcription_settings ) soundfont_bank.change( fn=lambda x: gr.update(visible=(x == SYNTH_8_BIT_LABEL)), inputs=soundfont_bank, outputs=synth_8bit_settings ) # --- Event listener for the new basic_pitch preset dropdown --- basic_pitch_preset_selector.change( fn=apply_basic_pitch_preset, inputs=basic_pitch_preset_selector, outputs=basic_pitch_ui_components ) # --- Event listener for the 8-bit preset selector --- s8bit_preset_selector.change( fn=apply_8bit_preset, inputs=s8bit_preset_selector, outputs=s8bit_control_components ) # --- New event listener for the render_type radio button --- # This listener now has TWO outputs render_type.change( fn=update_advanced_midi_options_visibility, inputs=render_type, outputs=advanced_rendering_options ).then( # Chain another event to the same trigger fn=update_render_type_description, inputs=render_type, outputs=render_type_info ) # Launch the Gradio app app.queue().launch(inbrowser=True, debug=True)