app.py

import torch
import torchaudio
import gradio as gr

from zonos.model import Zonos
from zonos.conditioning import make_cond_dict

model = Zonos.from_pretrained("Zyphra/Zonos-v0.1-hybrid", device="cuda")
model.bfloat16()

def tts(text, reference_audio):
    if reference_audio is None:
        return None
    
    # Gradio returns (sample_rate, audio_data) for type="numpy"
    sr, wav_np = reference_audio
    
    # Convert NumPy audio data to Torch tensor
    wav_torch = torch.from_numpy(wav_np).float().unsqueeze(0)
    if wav_torch.dim() == 2 and wav_torch.shape[0] > wav_torch.shape[1]:
        wav_torch = wav_torch.T

    # Create speaker embedding
    spk_embedding = model.embed_spk_audio(wav_torch, sr)
    
    # Prepare conditioning
    cond_dict = make_cond_dict(
        text=text,
        speaker=spk_embedding.to(torch.bfloat16),
        language="en-us",
    )
    conditioning = model.prepare_conditioning(cond_dict)
    
    # Generate codes & decode
    with torch.no_grad():
        torch.manual_seed(421)
        codes = model.generate(conditioning)
    
    wavs = model.autoencoder.decode(codes).cpu()
    out_audio = wavs[0].numpy()
    
    # Return a tuple of (sample_rate, audio_data) for playback
    return (model.autoencoder.sampling_rate, out_audio)

demo = gr.Interface(
    fn=tts,
    inputs=[
        gr.Textbox(label="Text to Synthesize"),
        gr.Audio(type="numpy", label="Reference Audio (Speaker)"),
    ],
    outputs=gr.Audio(label="Generated Audio"),
    title="Zonos TTS Demo (Hybrid)",
    description="Upload a reference audio for speaker embedding, enter text, and generate speech!"
)

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