app.py

import torch
import torchaudio
import gradio as gr

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

# We'll keep a global dictionary of loaded models to avoid reloading
MODELS_CACHE = {}
device = "cuda"

def load_model(model_name: str):
    """
    Loads or retrieves a cached Zonos model, sets it to eval and bfloat16.
    """
    global MODELS_CACHE
    if model_name not in MODELS_CACHE:
        print(f"Loading model: {model_name}")
        model = Zonos.from_pretrained(model_name, device=device)
        model = model.requires_grad_(False).eval()
        model.bfloat16()  # optional if GPU supports bfloat16
        MODELS_CACHE[model_name] = model
        print(f"Model loaded successfully: {model_name}")
    return MODELS_CACHE[model_name]

def tts(text, speaker_audio, selected_language, model_choice):
    """
    text: str (Text prompt to synthesize)
    speaker_audio: (sample_rate, numpy_array) from Gradio if type="numpy"
    selected_language: str (language code)
    model_choice: str (which Zonos model to use, e.g., "Zyphra/Zonos-v0.1-hybrid")
    
    Returns (sample_rate, waveform) for Gradio audio output.
    """
    # Load the selected model
    model = load_model(model_choice)

    if not text:
        return None
    if speaker_audio is None:
        return None

    # Gradio gives audio in the format (sample_rate, numpy_array)
    sr, wav_np = speaker_audio

    # Convert to Torch tensor: shape (1, num_samples)
    wav_tensor = torch.from_numpy(wav_np).unsqueeze(0).float()
    if wav_tensor.dim() == 2 and wav_tensor.shape[0] > wav_tensor.shape[1]:
        # If shape is transposed, fix it
        wav_tensor = wav_tensor.T

    # Get speaker embedding
    with torch.no_grad():
        spk_embedding = model.make_speaker_embedding(wav_tensor, sr)
        spk_embedding = spk_embedding.to(device, dtype=torch.bfloat16)

    # Prepare conditioning dictionary
    cond_dict = make_cond_dict(
        text=text,
        speaker=spk_embedding,
        language=selected_language,
        device=device,
    )
    conditioning = model.prepare_conditioning(cond_dict)

    # Generate codes
    with torch.no_grad():
        codes = model.generate(conditioning)

    # Decode the codes into raw audio
    wav_out = model.autoencoder.decode(codes).cpu().detach().squeeze()
    sr_out = model.autoencoder.sampling_rate

    return (sr_out, wav_out.numpy())

def build_demo():
    with gr.Blocks() as demo:
        gr.Markdown("# Simple Zonos TTS Demo")

        with gr.Row():
            text_input = gr.Textbox(
                label="Text Prompt",
                value="Hello from Zonos!",
                lines=3
            )
            ref_audio_input = gr.Audio(
                label="Reference Audio (Speaker Cloning)",
                type="numpy"
            )

        # Model dropdown
        model_dropdown = gr.Dropdown(
            label="Model Choice",
            choices=["Zyphra/Zonos-v0.1-transformer", "Zyphra/Zonos-v0.1-hybrid"],
            value="Zyphra/Zonos-v0.1-hybrid",
            interactive=True,
        )
        # Language dropdown (you can filter or use all from supported_language_codes)
        language_dropdown = gr.Dropdown(
            label="Language Code",
            choices=["en-us", "es-es", "fr-fr", "de-de", "it"],
            value="en-us",
            interactive=True,
        )

        generate_button = gr.Button("Generate")
        audio_output = gr.Audio(label="Synthesized Output", type="numpy")

        generate_button.click(
            fn=tts,
            inputs=[text_input, ref_audio_input, language_dropdown, model_dropdown],
            outputs=audio_output,
        )

    return demo

if __name__ == "__main__":
    demo_app = build_demo()
    demo_app.launch(server_name="0.0.0.0", server_port=7860, share=True)