app.py

import os
import tempfile
from TTS.api import TTS
import gradio as gr

# Global singleton: model loaded only once
tts = None
counter = 0  # To auto-number output files
MODEL_NAME = "tts_models/multilingual/multi-dataset/xtts_v2"

def load_model_once():
    global tts
    if tts is None:
        tts = TTS(model_name=MODEL_NAME, gpu=False)
    return tts

def generate_audio(text, speaker_wav):
    global counter
    if not text or speaker_wav is None:
        return None

    tts_model = load_model_once()
    
    counter += 1
    out_path = os.path.join(tempfile.gettempdir(), f"{counter}.wav")

    tts_model.tts_to_file(
        text=text,
        speaker_wav=speaker_wav,
        language="hi",
        file_path=out_path
    )

    return out_path

with gr.Blocks(title="XTTS Voice Clone (CPU)") as demo:
    gr.Markdown("## 🎙️ XTTS Voice Cloning (CPU Space)\nUpload voice + enter text → get WAV")

    with gr.Row():
        text_in = gr.Textbox(label="Enter Text (Hindi / English)", lines=4, placeholder="नमस्ते...")
        wav_in = gr.Audio(label="Upload Voice Sample (WAV)", type="filepath")

    out_audio = gr.Audio(label="Generated Voice")
    btn = gr.Button("Generate Voice")

    # ⚡ Set concurrency_limit directly instead of deprecated queue()
    btn.click(
        fn=generate_audio,
        inputs=[text_in, wav_in],
        outputs=out_audio,
        concurrency_limit=2  # adjust as needed
    )

demo.launch(max_threads=2)  # adjust max_threads if needed