app.py

import gradio as gr
import soundfile as sf
import tempfile
from voxcpm import VoxCPM

# ===============================
# Load VoxCPM model once
# ===============================
model = VoxCPM.from_pretrained("openbmb/VoxCPM-0.5B")

# ===============================
# TTS function
# ===============================
def tts_generate(text, cfg_value, inference_steps, normalize, denoise):
    try:
        wav = model.generate(
            text=text,
            cfg_value=cfg_value,
            inference_timesteps=inference_steps,
            normalize=normalize,
            denoise=denoise
        )

        tmp_wav = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
        sf.write(tmp_wav.name, wav, 16000)
        return tmp_wav.name
    except Exception as e:
        return f"Error: {str(e)}"

# ===============================
# Gradio Interface
# ===============================
app = gr.Interface(
    fn=tts_generate,
    inputs=[
        gr.Textbox(label="Enter text", value="Hello Hugging Face!", lines=3),
        gr.Slider(0.5, 5.0, value=2.0, step=0.1, label="CFG Value"),
        gr.Slider(5, 20, value=8, step=1, label="Inference timesteps"),
        gr.Checkbox(value=True, label="Enable Normalization"),
        gr.Checkbox(value=True, label="Enable Denoise"),
    ],
    outputs=gr.Audio(type="filepath", label="Generated Audio"),
    title="🎙️ VoxCPM Text-to-Speech",
    description="Generate expressive speech with VoxCPM TTS. Adjust CFG for text fidelity and inference timesteps for speed vs quality."
)

# ===============================
# Run the app
# ===============================
if __name__ == "__main__":
    app.launch()