app.py

import gradio as gr
import torch
import numpy as np
from transformers import VitsModel, AutoTokenizer

LANG_MODEL_MAP = {
    "Malayalam": "trysem/Seema",
}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cache = {}

def load_model_and_tokenizer(language):
    model_name = LANG_MODEL_MAP[language]
    if model_name not in cache:
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        model = VitsModel.from_pretrained(model_name).to(device)
        cache[model_name] = (tokenizer, model)
    return cache[model_name]

def tts(language, text):
    if not text.strip():
        return 16000, np.zeros(1)  # empty waveform if no text

    tokenizer, model = load_model_and_tokenizer(language)
    inputs = tokenizer(text, return_tensors="pt").to(device)

    with torch.no_grad():
        output = model(**inputs)

    waveform = output.waveform.squeeze().cpu().numpy()
    return 16000, waveform

iface = gr.Interface(
    fn=tts,
    inputs=[
        gr.Dropdown(choices=list(LANG_MODEL_MAP.keys()), label="Select Language"),
        gr.Textbox(label="Enter Text")
    ],
    outputs=gr.Audio(label="Synthesized Speech", type="numpy"),
    title="Multilingual Text-to-Speech (MMS)",
    description="Generate speech from text using Meta's MMS models for Malayalam"
)

if __name__ == "__main__":
    iface.launch()