app.py

from audiocraft.models import MusicGen
import streamlit as st
import os
import torch
import torchaudio
import numpy as np
import base64
from io import BytesIO

@st.cache_resource
def load_model():
    model = MusicGen.get_pretrained("facebook/musicgen-small")
    return model

def generate_music_tensors(description, duration: int):
    print("Description:", description)
    print("Duration:", duration)
    model = load_model()

    # Experiment with different generation parameters for improved quality
    model.set_generation_params(
        use_sampling=True,
        top_k=300,          # Increase top_k for more diversity
        top_p=0.85,         # Probability threshold for token sampling
        temperature=0.8,    # Control randomness; lower values = more focused output
        duration=duration
    )

    output = model.generate(
        descriptions=[description],
        progress=True,
        return_tokens=True
    )
    return output[0]

def save_audio_to_bytes(samples: torch.Tensor):
    sample_rate = 32000
    assert samples.dim() == 2 or samples.dim() == 3
    samples = samples.detach().cpu()

    if samples.dim() == 2:
        samples = samples[None, ...]
    
    # Save audio to a byte buffer instead of file for easier download
    byte_io = BytesIO()
    torchaudio.save(byte_io, samples, sample_rate=sample_rate, format="wav")
    byte_io.seek(0)  # Reset buffer position to the beginning for reading
    return byte_io

st.set_page_config(
    page_icon=":musical_note:",
    page_title="Music Gen"
)

def main():
    st.title("Your Music")

    with st.expander("See Explanation"):
        st.write("App is developed using Meta's Audiocraft Music Gen model. Write a description and we will generate audio.")

    text_area = st.text_area("Enter description")
    time_slider = st.slider("Select time duration (seconds)", 2, 20, 10)

    if text_area and time_slider:
        st.json(
            {
                "Description": text_area,
                "Selected duration": time_slider
            }
            st.write("We will back with your music....please enjoy doing the rest of your tasks while we come back in some time :)")
        )
        st.subheader("Generated Music")
        music_tensors = generate_music_tensors(text_area, time_slider)

        # Save to byte buffer for download
        audio_file = save_audio_to_bytes(music_tensors)

        # Play and download audio
        st.audio(audio_file, format="audio/wav")
        st.download_button(
            label="Download Audio",
            data=audio_file,
            file_name="generated_music.wav",
            mime="audio/wav"
        )

if __name__ == "__main__":
    main()