app.py

import os
import streamlit as st
import whisper
from transformers import pipeline
from gtts import gTTS
import speech_recognition as sr
import tempfile
from langchain_community.vectorstores import FAISS
from langchain.document_loaders import PyPDFLoader
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.chains import ConversationalRetrievalChain
from langchain.memory import ConversationBufferMemory

# Initialize models
whisper_model = whisper.load_model("base")  # Use the base model for faster performance
translation_pipeline = pipeline(
    "translation", model="Helsinki-NLP/opus-mt-ur-en-tiny", tokenizer="Helsinki-NLP/opus-mt-ur-en-tiny"
)
urdu_translation_pipeline = pipeline(
    "translation", model="Helsinki-NLP/opus-mt-en-ur-tiny", tokenizer="Helsinki-NLP/opus-mt-en-ur-tiny"
)
embedding_model = HuggingFaceEmbeddings(model_name="sentence-transformers/paraphrase-MiniLM-L6-v2")

# Streamlit interface
st.title("Real-Time Voice-to-Voice First Aid Chatbot")

uploaded_file = st.file_uploader("Upload a PDF file for First Aid Knowledge", type=["pdf"])
if uploaded_file:
    st.write("Processing PDF...")
    loader = PyPDFLoader(uploaded_file)
    documents = loader.load()
    
    st.write("Creating vector database...")
    vectorstore = FAISS.from_documents(documents, embedding_model)
    st.write("Knowledge base ready.")
    
    memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)
    chain = ConversationalRetrievalChain.from_llm(
        llm=None,  # Replace with a valid LLM integration like OpenAI or Groq client
        retriever=vectorstore.as_retriever(),
        memory=memory,
    )

if st.button("Start Chat"):
    st.write("Listening... Speak now!")
    recognizer = sr.Recognizer()
    
    with sr.Microphone() as source:
        st.write("Adjusting for ambient noise, please wait...")
        recognizer.adjust_for_ambient_noise(source)
        st.write("You can now speak.")
        
        while True:
            try:
                st.write("Listening...")
                audio = recognizer.listen(source)
                st.write("Processing audio...")
                
                with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_audio:
                    temp_audio.write(audio.get_wav_data())
                    temp_audio_path = temp_audio.name
                
                transcription = whisper_model.transcribe(temp_audio_path)["text"]
                st.write(f"You said: {transcription}")
                
                translated_text = translation_pipeline(transcription)[0]["translation_text"]
                st.write(f"Translated Text: {translated_text}")
                
                response = chain({"input": translated_text})["response"]
                st.write(f"Response: {response}")
                
                urdu_response = urdu_translation_pipeline(response)[0]["translation_text"]
                st.write(f"Response in Urdu: {urdu_response}")
                
                tts = gTTS(urdu_response, lang="ur")
                response_audio_path = tempfile.NamedTemporaryFile(delete=False, suffix=".mp3").name
                tts.save(response_audio_path)
                os.system(f"mpg123 {response_audio_path}")
            
            except Exception as e:
                st.write(f"Error: {str(e)}")