app.py

import gradio as gr
from faster_whisper import WhisperModel
from transformers import MarianMTModel, MarianTokenizer
from TTS.api import TTS
import os

# Load Whisper for Hindi ASR
whisper_model = WhisperModel("medium", compute_type="float32", download_root="./models")

# Load Helsinki-NLP Hindi-to-English Translator
translator_name = "Helsinki-NLP/opus-mt-hi-en"
translator_tokenizer = MarianTokenizer.from_pretrained(translator_name)
translator_model = MarianMTModel.from_pretrained(translator_name)

# Load TTS model for English voice
tts = TTS(model_name="tts_models/en/ljspeech/tacotron2-DDC", progress_bar=True, gpu=False)

def translate_hi_to_en(text):
    inputs = translator_tokenizer(text, return_tensors="pt", padding=True, truncation=True)
    translated = translator_model.generate(**inputs)
    return translator_tokenizer.decode(translated[0], skip_special_tokens=True)

def transcribe_and_translate(audio_path):
    # Step 1: Transcribe Hindi audio
    segments, _ = whisper_model.transcribe(audio_path, language="hi")
    hindi_text = " ".join([segment.text for segment in segments])

    # Step 2: Translate to English
    english_text = translate_hi_to_en(hindi_text)

    # Step 3: Convert English text to speech
    output_audio_path = "output.wav"
    tts.tts_to_file(text=english_text, file_path=output_audio_path)

    return english_text, output_audio_path

# Gradio Interface
iface = gr.Interface(
    fn=transcribe_and_translate,
    inputs=gr.Audio(type="filepath", label="Speak in Hindi"),
    outputs=[
        gr.Textbox(label="Translated English Text"),
        gr.Audio(type="filepath", label="English Speech")
    ],
    title="Hindi to English Speech Translator",
    description="🎤 Speak Hindi → 📘 Translate to English → 🔊 English Speech"
)

iface.launch()