leenag/Multilingual_TTS

1import gradio as gr2import torch3import numpy as np4from transformers import VitsModel, AutoTokenizer5 6LANG_MODEL_MAP = {7    "English": "facebook/mms-tts-eng",8    "Hindi": "facebook/mms-tts-hin",9    "Tamil": "facebook/mms-tts-tam",10    "Malayalam": "facebook/mms-tts-mal",11    "Kannada": "facebook/mms-tts-kan",12    "Telugu": "facebook/mms-tts-tel"13}14 15device = torch.device("cuda" if torch.cuda.is_available() else "cpu")16cache = {}17 18def load_model_and_tokenizer(language):19    model_name = LANG_MODEL_MAP[language]20    if model_name not in cache:21        tokenizer = AutoTokenizer.from_pretrained(model_name)22        model = VitsModel.from_pretrained(model_name).to(device)23        cache[model_name] = (tokenizer, model)24    return cache[model_name]25 26def tts(language, text):27    if not text.strip():28        return 16000, np.zeros(1)  # empty waveform if no text29 30    tokenizer, model = load_model_and_tokenizer(language)31    inputs = tokenizer(text, return_tensors="pt").to(device)32 33    with torch.no_grad():34        output = model(**inputs)35 36    waveform = output.waveform.squeeze().cpu().numpy()37    return 16000, waveform38 39iface = gr.Interface(40    fn=tts,41    inputs=[42        gr.Dropdown(choices=list(LANG_MODEL_MAP.keys()), label="Select Language"),43        gr.Textbox(label="Enter Text")44    ],45    outputs=gr.Audio(label="Synthesized Speech", type="numpy"),46    title="Multilingual Text-to-Speech (MMS)",47    description="Generate speech from text using Meta's MMS models for English, Hindi, Tamil, Malayalam, Kannada and Telugu."48)49 50if __name__ == "__main__":51    iface.launch()