app.py

import gradio as gr
import numpy as np
import soundfile as sf
import os
import uuid

import torch
from transformers import VitsModel, VitsTokenizer, set_seed

# 1. Load MMS-TTS English model (lighter than Bark)
MODEL_ID = "facebook/mms-tts-eng"

tokenizer = VitsTokenizer.from_pretrained(MODEL_ID)
model = VitsModel.from_pretrained(MODEL_ID)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# Optional: make outputs deterministic
set_seed(555)


MAX_CHARS = 300  # keep text short for speed and stability


def generate_speech(text: str) -> str:
    """
    Take text, synthesize speech with MMS-TTS,
    save to a WAV file, and return the filepath
    (for gr.Audio(type="filepath")).
    """
    if not text or text.strip() == "":
        raise gr.Error("Please enter some text 🙂")

    text = text.strip()
    if len(text) > MAX_CHARS:
        text = text[:MAX_CHARS]
        # You could also show a warning text if you like.

    # MMS-TTS is trained on lowercased, unpunctuated text → simple normalization
    normalized_text = text.lower()

    # 1) Tokenize
    inputs = tokenizer(text=normalized_text, return_tensors="pt").to(device)

    # 2) Forward pass
    with torch.no_grad():
        outputs = model(**inputs)

    # 3) Get waveform and sampling rate
    waveform = outputs.waveform[0].cpu().numpy().astype(np.float32)
    sr = model.config.sampling_rate  # typically 16000

    # 4) Save to /tmp as WAV
    tmp_dir = "/tmp"
    os.makedirs(tmp_dir, exist_ok=True)
    filename = f"tts_{uuid.uuid4().hex}.wav"
    filepath = os.path.join(tmp_dir, filename)

    sf.write(filepath, waveform, sr)

    # 5) Return file path for gr.Audio(type="filepath")
    return filepath


with gr.Blocks() as demo:
    gr.Markdown("# 🗣️ Англи текстийг яриа болгох \n\n --- Simple TTS with facebook/mms-tts-eng")
    gr.Markdown(
        "Энд англи дээр өгүүлбэрээ бичээд **Яриаг үүсгэ** товчийг дарж англи яриаг сонсоорой. \n\n"
        "Model: `facebook/mms-tts-eng` (MMS-TTS, VITS-based)."
    )

    with gr.Row():
        with gr.Column(scale=2):
            text_input = gr.Textbox(
                label="Яриа болгох англи өгүүлбэр",
                placeholder="Жишээ: Hello, this is my text-to-speech demo",
                lines=3,
            )
            generate_button = gr.Button("Яриаг үүсгэнэ үү", variant="primary")
        with gr.Column(scale=1):
            audio_output = gr.Audio(
                label="Үүссэн бичлэг",
                type="filepath",  # we return a path string
            )

    generate_button.click(
        fn=generate_speech,
        inputs=text_input,
        outputs=audio_output,
    )

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