app.py

import gradio as gr
import torch
import torchaudio
from transformers import MusicgenForConditionalGeneration, MusicgenProcessor

model = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-melody", torch_dtype=torch.float32)
processor = MusicgenProcessor.from_pretrained("facebook/musicgen-melody")

def generate_music(prompt, melody):
    if melody is None:
        return None

    # Load and resample melody to 32kHz
    melody_waveform, melody_sr = torchaudio.load(melody)
    if melody_sr != 32000:
        resampler = torchaudio.transforms.Resample(orig_freq=melody_sr, new_freq=32000)
        melody_waveform = resampler(melody_waveform)

    # Trim or pad to 30 seconds
    melody_waveform = melody_waveform[:, :32000 * 30]

    # Run the model
    inputs = processor(audio=melody_waveform, sampling_rate=32000, text=[prompt], return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=1024)

    audio_array = outputs[0].cpu().numpy()
    return (audio_array, model.config.audio_encoder.sampling_rate)

demo = gr.Interface(
    fn=generate_music,
    inputs=[
        gr.Textbox(label="Prompt", placeholder="e.g., mellow lofi beat with piano"),
        gr.Audio(source="upload", type="filepath", label="Melody Input (WAV or MP3)")
    ],
    outputs=gr.Audio(label="Generated Track"),
    title="🎵 MusicGen-Melody AI Generator",
    description="Upload a melody and describe the vibe. Generates music using Meta’s MusicGen-Melody model."
)

demo.launch()