app.py

import streamlit as st
from transformers import MarianMTModel, MarianTokenizer, pipeline
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech
import soundfile as sf
import torch

# Define the language model and tokenizer
translation_model_name = "Helsinki-NLP/opus-mt-en-ur"
tokenizer = MarianTokenizer.from_pretrained(translation_model_name)
translation_model = MarianMTModel.from_pretrained(translation_model_name)

# Load the text-to-speech model
tts_model_name = "microsoft/speecht5_tts"
processor = SpeechT5Processor.from_pretrained(tts_model_name)
tts_model = SpeechT5ForTextToSpeech.from_pretrained(tts_model_name)

# Function to translate text
def translate_text(text, target_lang):
    if target_lang not in ["Urdu", "Hindi", "Bengali"]:
        return "Error: Target language not supported."
    tokens = tokenizer(text, return_tensors="pt", padding=True)
    translated_tokens = translation_model.generate(**tokens)
    translated_text = tokenizer.decode(translated_tokens[0], skip_special_tokens=True)
    return translated_text

# Function to generate speech
def synthesize_speech(text, speaker_embeddings=None):
    inputs = processor(text, return_tensors="pt")
    with torch.no_grad():
        speech = tts_model.generate_speech(inputs, speaker_embeddings)
    sf.write("output.wav", speech.numpy(), 16000)
    return "output.wav"

# Streamlit app
st.title("Language Translator with TTS")
st.write("Enter the text you want to translate and hear the translation.")

# Select target language
target_language = st.selectbox("Select Target Language", ["Urdu", "Hindi", "Bengali"])

# Text input
text_to_translate = st.text_input("Enter text here")

if st.button("Translate and Generate Audio"):
    # Clear input for new text
    st.session_state.text_to_translate = ""

    # Perform translation
    translated_text = translate_text(text_to_translate, target_language)
    st.write(f"Translated text ({target_language}): {translated_text}")

    # Generate speech
    if translated_text and "Error" not in translated_text:
        speaker_embeddings = None  # Placeholder, use actual embeddings if needed
        audio_file = synthesize_speech(translated_text, speaker_embeddings)
        st.audio(audio_file)