import os import gradio as gr import ctranslate2 from transformers import AutoTokenizer LANG_TOKEN = {"assamese": "<2asm>", "hindi": "<2hin>", "marathi": "<2mar>"} MODEL_DIR = "ct2_mt5_xlit" # directory in Space repo TOKENIZER_DIR = "tokenizer" # directory in Space repo tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_DIR) device = "cuda" if ctranslate2.get_cuda_device_count() > 0 else "cpu" compute_type = "float16" if device == "cuda" else "int8" translator = ctranslate2.Translator(MODEL_DIR, device=device, compute_type=compute_type) def transliterate(text, lang): lang = lang.lower() tagged = f"{LANG_TOKEN[lang]} {text.strip().lower()}" enc = tokenizer([tagged], return_tensors=None, padding=True, truncation=True, max_length=24) tokens = [tokenizer.convert_ids_to_tokens(enc["input_ids"][0])] out = translator.translate_batch(tokens, beam_size=3, max_decoding_length=24)[0] return tokenizer.convert_tokens_to_string(out.hypotheses[0]).strip() demo = gr.Interface( fn=transliterate, inputs=[ gr.Textbox(label="Input (romanized)", placeholder="e.g., Saurav"), gr.Dropdown(["hindi", "marathi", "assamese"], value="hin", label="Target language"), ], outputs=gr.Textbox(label="Transliteration"), title="Multilingual Transliteration", ) if __name__ == "__main__": demo.launch()