app.py

# app.py
import gradio as gr
import torch, torchaudio, spaces
from zonos.model import Zonos
from zonos.conditioning import make_cond_dict

cpu = "cpu"
model = Zonos.from_pretrained("Zyphra/Zonos-v0.1-hybrid", device=cpu)

def _speaker_embed(audio):
    if audio is None:
        return None
    sr, wav = audio
    wav = torch.tensor(wav).unsqueeze(0)
    return model.make_speaker_embedding(wav, sr)

@spaces.GPU
def tts(text, language, speaker_audio, emotion_vec, speaking_rate, pitch_std):
    speaker = _speaker_embed(speaker_audio)
    emotion = [float(x) for x in emotion_vec.split(",")] if emotion_vec else None
    model.to("cuda")
    with torch.no_grad():
        cond = make_cond_dict(
            text=text,
            language=language,
            speaker=speaker,
            emotion=emotion,
            speaking_rate=float(speaking_rate),
            pitch_std=float(pitch_std),
        )
        conditioning = model.prepare_conditioning(cond)
        codes = model.generate(conditioning)
        wav = model.autoencoder.decode(codes)[0].cpu()
    model.to(cpu)
    torch.cuda.empty_cache()
    return (model.autoencoder.sampling_rate, wav.numpy())

langs = ["en-us", "fr-fr", "ja", "de-de", "zh"]
demo = gr.Interface(
    fn=tts,
    inputs=[
        gr.Textbox(label="text"),
        gr.Dropdown(langs, value="en-us", label="language"),
        gr.Audio(type="numpy", label="speaker reference (optional)"),
        gr.Textbox(value="0.3,0,0,0,0,0,0.2,0.5", label="emotion (8 comma-sep floats)"),
        gr.Slider(0, 40, value=15, step=1, label="speaking_rate"),
        gr.Slider(0, 400, value=20, step=1, label="pitch_std"),
    ],
    outputs=gr.Audio(label="generated speech"),
    title="zonos-v0.1 zerogpu tts",
)

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