app.py

import gradio as gr
from omnivoice import OmniVoice
import soundfile as sf
import torch
import os
import uuid

# -----------------------------
# Load model (CPU version)
# -----------------------------
print("Loading OmniVoice model on CPU...")

model = OmniVoice.from_pretrained(
    "k2-fsa/OmniVoice",
    device_map="cpu",          # IMPORTANT: CPU only
    dtype=torch.float32,       # IMPORTANT: float32 for CPU
    load_asr=False,
)

print("Model loaded successfully!")

# -----------------------------
# Inference function
# -----------------------------
def clone_voice(text, ref_audio):
    if ref_audio is None:
        return None

    try:
        # Generate audio
        audio = model.generate(
            text=text,
            ref_audio=ref_audio,
        )

        # Save output file
        output_path = f"output_{uuid.uuid4().hex}.wav"
        sf.write(output_path, audio[0], 24000)

        return output_path

    except Exception as e:
        print("Error:", str(e))
        return None


# -----------------------------
# Gradio UI
# -----------------------------
iface = gr.Interface(
    fn=clone_voice,
    inputs=[
        gr.Textbox(label="Text", value="Hello, this is a test of zero-shot voice cloning."),
        gr.Audio(type="filepath", label="Reference Audio"),
    ],
    outputs=gr.Audio(type="filepath", label="Cloned Audio"),
    title="OmniVoice Voice Cloning API",
    description="Upload a voice sample and generate cloned speech.",
)

# -----------------------------
# Launch (important for Spaces)
# -----------------------------
if __name__ == "__main__":
    iface.launch(server_name="0.0.0.0", server_port=7860)