########################################## # Step 0: Import required libraries ########################################## import streamlit as st # For building the web application interface from transformers import ( # For text classification, text-to-speech, and text generation pipeline, SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan, AutoModelForCausalLM, AutoTokenizer ) from datasets import load_dataset # To load speaker embeddings dataset import torch # For tensor operations import soundfile as sf # For saving audio as .wav files import sentencepiece # Required by SpeechT5Processor for tokenization ########################################## # Streamlit application title and input ########################################## # Display a deep blue title in a large, visually appealing font st.markdown( "

🚀 Just Comment

", unsafe_allow_html=True ) # Set deep blue title # Display a gentle, warm subtitle below the title st.markdown( "

I'm listening to you, my friend～

", unsafe_allow_html=True ) # Set a friendly subtitle # Add a text area for user input with placeholder and tooltip text = st.text_area( "Enter your comment", placeholder="Type something here...", height=100, help="Write a comment you would like us to respond to!" # Provide tooltip ) # Create text input field ########################################## # Step 1: Sentiment Analysis Function ########################################## def analyze_dominant_emotion(user_review): """ Analyze the dominant emotion in the user's comment using a fine-tuned text classification model. """ emotion_classifier = pipeline( "text-classification", model="Thea231/jhartmann_emotion_finetuning", return_all_scores=True ) # Load the sentiment classification model emotion_results = emotion_classifier(user_review)[0] # Get sentiment scores for the input text dominant_emotion = max(emotion_results, key=lambda x: x['score']) # Identify the emotion with highest score return dominant_emotion # Return the dominant emotion (as a dict with label and score) ########################################## # Step 2: Response Generation Functions ########################################## def prompt_gen(user_review): """ Generate the text generation prompt based on the user's comment and detected emotion. """ # Get dominant emotion for the input dominant_emotion = analyze_dominant_emotion(user_review) # Analyze user's comment # Define response templates for 7 emotions emotion_strategies = { "anger": { "prompt": ( "Customer complaint: '{review}'\n\n" "As a customer service representative, craft a professional response that:\n" "- Begins with a sincere apology and acknowledgment.\n" "- Clearly explains a solution process with concrete steps.\n" "- Offers appropriate compensation or redemption.\n" "- Keeps a humble and solution-focused tone (1-3 sentences).\n\n" "Response:" ) }, "disgust": { "prompt": ( "Customer quality concern: '{review}'\n\n" "As a customer service representative, craft a response that:\n" "- Immediately acknowledges the product issue.\n" "- Explains measures taken in quality control.\n" "- Provides clear return/replacement instructions.\n" "- Offers a goodwill gesture (1-3 sentences).\n\n" "Response:" ) }, "fear": { "prompt": ( "Customer safety concern: '{review}'\n\n" "As a customer service representative, craft a reassuring response that:\n" "- Directly addresses the safety worries.\n" "- References relevant certifications or standards.\n" "- Offers a dedicated support contact.\n" "- Provides a satisfaction guarantee (1-3 sentences).\n\n" "Response:" ) }, "joy": { "prompt": ( "Customer review: '{review}'\n\n" "As a customer service representative, craft a concise response that:\n" "- Thanks the customer for their feedback.\n" "- Acknowledges both positive and constructive points.\n" "- Invites them to explore loyalty or referral programs (1-3 sentences).\n\n" "Response:" ) }, "neutral": { "prompt": ( "Customer feedback: '{review}'\n\n" "As a customer service representative, craft a balanced response that:\n" "- Provides additional relevant product information.\n" "- Highlights key service features.\n" "- Politely requests more detailed feedback.\n" "- Maintains a professional tone (1-3 sentences).\n\n" "Response:" ) }, "sadness": { "prompt": ( "Customer disappointment: '{review}'\n\n" "As a customer service representative, craft an empathetic response that:\n" "- Shows genuine understanding of the issue.\n" "- Proposes a personalized recovery solution.\n" "- Offers extended support options.\n" "- Maintains a positive outlook (1-3 sentences).\n\n" "Response:" ) }, "surprise": { "prompt": ( "Customer enthusiastic feedback: '{review}'\n\n" "As a customer service representative, craft a response that:\n" "- Matches the customer's positive energy.\n" "- Highlights unexpected product benefits.\n" "- Invites the customer to join community events or programs.\n" "- Maintains the brand's voice (1-3 sentences).\n\n" "Response:" ) } } # Dictionary mapping each emotion to a prompt template # Get the template for the detected emotion, default to 'neutral' if not found template = emotion_strategies.get(dominant_emotion["label"].lower(), emotion_strategies["neutral"])["prompt"] prompt = template.format(review=user_review) # Insert the user review into the template print(f"Generated prompt: {prompt}") # Debug print using f-string return prompt # Return the generated prompt def response_gen(user_review): """ Generate a response using text generation based on the user's comment. """ prompt = prompt_gen(user_review) # Get the generated prompt using the detected emotion template # Load the tokenizer and language model for text generation tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen1.5-0.5B") # Load tokenizer for text generation model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen1.5-0.5B") # Load causal language model for generation inputs = tokenizer(prompt, return_tensors="pt") # Tokenize the prompt outputs = model.generate( **inputs, max_new_tokens=100, # Allow up to 100 new tokens for the answer min_length=30, # Ensure a minimum length for the generated response no_repeat_ngram_size=2, # Avoid repeated phrases temperature=0.7 # Use a moderate temperature for creativity ) # Generate response from the model input_length = inputs.input_ids.shape[1] # Determine length of the input prompt response = tokenizer.decode(outputs[0][input_length:], skip_special_tokens=True) # Extract only generated answer text print(f"Generated response: {response}") # Debug print using f-string return response # Return the generated response ########################################## # Step 3: Text-to-Speech Conversion Function ########################################## def sound_gen(response): """ Convert the generated response to speech and embed an auto-playing audio player. """ # Load SpeechT5 processor, TTS model, and vocoder processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts") # Load TTS processor model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts") # Load TTS model vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan") # Load vocoder # Process the full generated response text for TTS inputs = processor(text=response, return_tensors="pt") # Convert text to model input tensors # Use dummy speaker embeddings with shape (1,768) to avoid dimension mismatch speaker_embeddings = torch.zeros(1, 768, dtype=torch.float32) # Create dummy speaker embedding spectrogram = model.generate_speech(inputs["input_ids"], speaker_embeddings) # Generate speech spectrogram with torch.no_grad(): speech = vocoder(spectrogram) # Convert spectrogram to waveform # Save the audio as a .wav file with 16kHz sampling rate sf.write("customer_service_response.wav", speech.numpy(), samplerate=16000) # Write the waveform to file st.audio("customer_service_response.wav", start_time=0) # Embed an auto-playing audio widget ########################################## # Main Function ########################################## def main(): """ Main function to orchestrate text generation and text-to-speech conversion. It displays only the generated response and plays its audio. """ if text: # Check if the user has entered a comment response = response_gen(text) # Generate a response using text generation based on emotion st.markdown( f"

{response}

", unsafe_allow_html=True ) # Display the generated response in styled format sound_gen(response) # Convert the generated response to speech and embed the audio player print(f"Final generated response: {response}") # Debug print using f-string # Execute the main function when the script is run if __name__ == "__main__": main() # Call the main function