app.py

##########################################
# Step 0: Import required libraries
##########################################
import streamlit as st  # For building the web application
from transformers import (
    pipeline,
    SpeechT5Processor,
    SpeechT5ForTextToSpeech,
    SpeechT5HifiGan,
    AutoModelForCausalLM,
    AutoTokenizer
)  # For emotion analysis, text-to-speech, and text generation
from datasets import load_dataset  # For loading datasets (e.g., speaker embeddings)
import torch  # For tensor operations
import soundfile as sf  # For saving audio as .wav files

##########################################
# Streamlit application title and input
##########################################
st.title("Comment Reply for You")  # Application title
st.write("Generate automatic replies for user comments")  # Application description
text = st.text_area("Enter your comment", "")  # Text input for user to enter comments

##########################################
# Step 1: Sentiment Analysis Function
##########################################
def analyze_dominant_emotion(user_review):
    """
    Analyze the dominant emotion in the user's review using a text classification model.
    """
    emotion_classifier = pipeline(
        "text-classification", 
        model="Thea231/jhartmann_emotion_finetuning", 
        return_all_scores=True
    )  # Load pre-trained emotion classification model
    
    emotion_results = emotion_classifier(user_review)[0]  # Get emotion scores for the review
    dominant_emotion = max(emotion_results, key=lambda x: x['score'])  # Find the emotion with the highest confidence
    return dominant_emotion

##########################################
# Step 2: Response Generation Function
##########################################
def response_gen(user_review):
    """
    Generate a response based on the sentiment of the user's review.
    """
    # Use Llama-based model to create a response based on a generated prompt
    dominant_emotion = analyze_dominant_emotion(user_review)  # Get the dominant emotion
    emotion_label = dominant_emotion['label'].lower()  # Extract emotion label
    
    # Define response templates for each emotion
    emotion_prompts = {
        "anger": (
            "Customer complaint: '{review}'\n\n"
            "As a customer service representative, write a response that:\n"
            "- Sincerely apologizes for the issue\n"
            "- Explains how the issue will be resolved\n"
            "- Offers compensation where appropriate\n\n"
            "Response:"
        ),
        "joy": (
            "Customer review: '{review}'\n\n"
            "As a customer service representative, write a positive response that:\n"
            "- Thanks the customer for their feedback\n"
            "- Acknowledges both positive and constructive comments\n"
            "- Invites them to explore loyalty programs\n\n"
            "Response:"
        ),
        # Add other emotions as needed...
    }
    
    # Format the prompt with the user's review
    prompt = emotion_prompts.get(emotion_label, "Neutral").format(review=user_review)
    
    # Load model directly
    from transformers import AutoTokenizer, AutoModelForCausalLM

    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B")
    model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-0.5B")
    
    inputs = tokenizer(prompt, return_tensors="pt")  # Tokenize the prompt
    outputs = model.generate(**inputs, max_new_tokens=100)  # Generate a response
    
    input_length = inputs.input_ids.shape[1]  # Length of the input text
    response = tokenizer.decode(outputs[0][input_length:], skip_special_tokens=True)  # Decode the generated text
    return response

##########################################
# Step 3: Text-to-Speech Conversion Function
##########################################
def sound_gen(response):
    """
    Convert the generated response to speech and save as a .wav file.
    """
    # Load the pre-trained TTS models
    processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
    model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
    vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
    
    # Load speaker embeddings (e.g., neutral female voice)
    embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
    speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
    
    # Process the input text and generate a spectrogram
    inputs = processor(text=response, return_tensors="pt")
    spectrogram = model.generate_speech(inputs["input_ids"], speaker_embeddings)
    
    # Use the vocoder to generate a waveform
    with torch.no_grad():
        speech = vocoder(spectrogram)
    
    # Save the generated speech as a .wav file
    sf.write("customer_service_response.wav", speech.numpy(), samplerate=16000)
    st.audio("customer_service_response.wav")  # Play the audio in Streamlit

##########################################
# Main Function
##########################################
def main():
    """
    Main function to orchestrate the workflow of sentiment analysis, response generation, and text-to-speech.
    """
    if text:  # Check if the user entered a comment
        response = response_gen(text)  # Generate a response
        st.write(f"Generated response: {response}")  # Display the generated response
        sound_gen(response)  # Convert the response to speech and play it

# Run the main function
if __name__ == "__main__":
    main()