app.py

import os
import time
import tempfile

import gradio as gr
import soundfile as sf
import torch
from qwen_tts import Qwen3TTSModel


MODEL_ID = os.getenv("MODEL_ID", "Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice")

model = None


def get_model():
    global model

    if model is not None:
        return model

    if torch.cuda.is_available():
        model = Qwen3TTSModel.from_pretrained(
            MODEL_ID,
            device_map="cuda:0",
            dtype=torch.bfloat16,
        )
    else:
        model = Qwen3TTSModel.from_pretrained(
            MODEL_ID,
            device_map="cpu",
            dtype=torch.float32,
        )

    return model


def generate_tts(text, language, speaker, instruction):
    text = (text or "").strip()
    instruction = (instruction or "").strip()

    if not text:
        raise gr.Error("Écris une phrase à synthétiser.")

    tts = get_model()

    wavs, sr = tts.generate_custom_voice(
        text=text,
        language=language,
        speaker=speaker,
        instruct=instruction,
    )

    output_path = os.path.join(
        tempfile.gettempdir(),
        f"qwen_tts_{int(time.time() * 1000)}.wav",
    )

    sf.write(output_path, wavs[0], sr)

    intent_json = {
        "detected_glosses": [],
        "detected_facial_expression": "not_connected_yet",
        "subtitle": text,
        "voice_instruction": instruction,
        "language": language,
        "speaker": speaker,
        "pipeline_stage": "tts_only_mvp",
    }

    return output_path, text, intent_json


with gr.Blocks(title="ASL to TTS MVP") as demo:
    gr.Markdown(
        """
        # ASL to TTS MVP

        Première version: on teste seulement la brique TTS.

        Ensuite, on branchera:
        video ASL -> glosses -> emotion -> intent JSON -> subtitle -> voice instruction -> TTS.
        """
    )

    with gr.Row():
        with gr.Column():
            text_input = gr.Textbox(
                label="Subtitle temporaire",
                value="Hello, I am happy to see you today.",
                lines=3,
            )

            instruction_input = gr.Textbox(
                label="Voice instruction",
                value="Speak with a warm, happy, expressive voice.",
                lines=2,
            )

            language_input = gr.Dropdown(
                label="Language",
                choices=[
                    "Auto",
                    "Chinese",
                    "English",
                    "Japanese",
                    "Korean",
                    "German",
                    "French",
                    "Russian",
                    "Portuguese",
                    "Spanish",
                    "Italian",
                ],
                value="English",
            )

            speaker_input = gr.Dropdown(
                label="Speaker",
                choices=[
                    "Vivian",
                    "Serena",
                    "Uncle_Fu",
                    "Dylan",
                    "Eric",
                    "Ryan",
                    "Aiden",
                    "Ono_Anna",
                    "Sohee",
                ],
                value="Ryan",
            )

            button = gr.Button("Generate speech")

        with gr.Column():
            audio_output = gr.Audio(label="Generated audio", type="filepath")
            subtitle_output = gr.Textbox(label="Subtitle")
            json_output = gr.JSON(label="Intent JSON")

    button.click(
        fn=generate_tts,
        inputs=[
            text_input,
            language_input,
            speaker_input,
            instruction_input,
        ],
        outputs=[
            audio_output,
            subtitle_output,
            json_output,
        ],
    )


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