app.py

import gradio as gr
from transformers import AutoModel
from miditok import MusicTokenizer
import torch

import numpy as np
import pretty_midi
from io import BytesIO
import os
import requests

# URL of the file to download
url = "https://raw.githubusercontent.com/urish/cinto/refs/heads/master/media/FluidR3%20GM.sf2"

# Local filename to save the file
filename = "FluidR3_GM.sf2"

if not os.path.exists(filename):
    response = requests.get(url)
    response.raise_for_status()

    with open(filename, "wb") as file:
        file.write(response.content)


def score_to_audio(score, sample_rate: int = 44100) -> tuple[int, np.ndarray]:
    """
    Convert a symusic Score to integer PCM audio data.

    Args:
        score (Score): symusic Score object
        sample_rate (int): desired sample rate in Hz, defaults to 44100

    Returns:
        tuple[int, np.ndarray]: Tuple of (sample_rate, audio_data as int16)
    """
    # Get MIDI bytes and create MIDI object
    midi_data = BytesIO(score.dumps_midi())
    pm = pretty_midi.PrettyMIDI(midi_data)

    # Synthesize to float array first
    # float_audio = pm.synthesize(fs=sample_rate)
    float_audio = pm.fluidsynth(
        fs=sample_rate, sf2_path="./FluidR3_GM.sf2"
    )

    # Convert to 16-bit integer PCM
    # Scale to full int16 range (-32768 to 32767)
    int_audio = (float_audio * 32767).astype(np.int16)
    int_audio = np.trim_zeros(int_audio, "fb")
    return sample_rate, int_audio


device = "cuda" if torch.cuda.is_available() else "cpu"
device = "cpu"

tokenizer = MusicTokenizer.from_pretrained("shikhr/music_maker")

model = AutoModel.from_pretrained("shikhr/music_maker", trust_remote_code=True)
model.to(device)


def generate_music():
    # Generate some music
    out = model.generate(
        torch.tensor([[1]]).to(device), max_new_tokens=200, temperature=1.0, top_k=100
    )

    # Save the generated MIDI
    res = tokenizer(out[0].tolist())
    res.dump_midi("output.mid")
    nx = score_to_audio(res)
    # print(nx)
    return "Generated" , nx


demo = gr.Interface(generate_music, inputs=[], outputs=["text", "audio"], flagging_mode="never")

if __name__ == "__main__":
    demo.launch()