app.py

import streamlit as st
import torch
import os
import librosa
import librosa.display
import matplotlib.pyplot as plt
from audiosr import build_model, super_resolution, save_wave
import tempfile
import numpy as np

# Set MPS device if available (for Mac M-Series GPUs)
device = "cuda" if torch.cuda.is_available() else "cpu"

# Title and Description
st.title("AudioSR: Versatile Audio Super-Resolution")
st.write("""
Upload your low-resolution audio files, and AudioSR will enhance them to high fidelity!
Supports all types of audio (music, speech, sound effects, etc.) with arbitrary sampling rates.
Only the first 10 seconds of the audio will be processed.
""")

# Upload audio file
uploaded_file = st.file_uploader("Upload an audio file (WAV format)", type=["wav"])

# Model Parameters
st.sidebar.title("Model Parameters")
model_name = st.sidebar.selectbox("Select Model", ["basic", "speech"], index=0)
ddim_steps = st.sidebar.slider("DDIM Steps", min_value=10, max_value=100, value=50)
guidance_scale = st.sidebar.slider("Guidance Scale", min_value=1.0, max_value=10.0, value=3.5)
random_seed = st.sidebar.number_input("Random Seed", min_value=0, value=42, step=1)
latent_t_per_second = 12.8

# Helper function to plot spectrogram
def plot_spectrogram(audio_path, title):
    y, sr = librosa.load(audio_path, sr=None)
    S = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128, fmax=sr // 2)
    S_dB = librosa.power_to_db(S, ref=np.max)
    
    plt.figure(figsize=(10, 4))
    librosa.display.specshow(S_dB, sr=sr, x_axis='time', y_axis='mel', fmax=sr // 2, cmap='viridis')
    plt.colorbar(format='%+2.0f dB')
    plt.title(title)
    plt.tight_layout()
    return plt

# Process Button
if uploaded_file and st.button("Enhance Audio"):
    st.write("Processing audio...")

    # Create temp directory for saving files
    with tempfile.TemporaryDirectory() as tmp_dir:
        input_path = os.path.join(tmp_dir, "input.wav")
        truncated_path = os.path.join(tmp_dir, "truncated.wav")
        output_path = os.path.join(tmp_dir, "output.wav")

        # Save uploaded file locally
        with open(input_path, "wb") as f:
            f.write(uploaded_file.read())

        # Load and truncate the first 10 seconds
        y, sr = librosa.load(input_path, sr=None)
        max_samples = sr * 10  # First 10 seconds
        y_truncated = y[:max_samples]
        librosa.output.write_wav(truncated_path, y_truncated, sr)

        # Plot truncated spectrogram
        st.write("Truncated Input Audio Spectrogram (First 10 seconds):")
        truncated_spectrogram = plot_spectrogram(truncated_path, title="Truncated Input Audio Spectrogram")
        st.pyplot(truncated_spectrogram)

        # Build and load the model
        audiosr = build_model(model_name=model_name, device=device)

        # Perform super-resolution
        waveform = super_resolution(
            audiosr,
            truncated_path,
            seed=random_seed,
            guidance_scale=guidance_scale,
            ddim_steps=ddim_steps,
            latent_t_per_second=latent_t_per_second,
        )

        # Save enhanced audio
        save_wave(waveform, inputpath=truncated_path, savepath=tmp_dir, name="output", samplerate=48000)

        # Plot output spectrogram
        st.write("Enhanced Audio Spectrogram:")
        output_spectrogram = plot_spectrogram(output_path, title="Enhanced Audio Spectrogram")
        st.pyplot(output_spectrogram)

        # Display audio players and download link
        st.audio(truncated_path, format="audio/wav")
        st.write("Truncated Original Audio (First 10 seconds):")
        
        st.audio(output_path, format="audio/wav")
        st.write("Enhanced Audio:")
        st.download_button("Download Enhanced Audio", data=open(output_path, "rb").read(), file_name="enhanced_audio.wav")

# Footer
st.write("Built with [Streamlit](https://streamlit.io) and [AudioSR](https://audioldm.github.io/audiosr)")