Spaces:
Running
Running
| import gradio as gr | |
| from llama_cpp import Llama | |
| import whisper | |
| from gtts import gTTS | |
| import tempfile | |
| import os | |
| from huggingface_hub import hf_hub_download | |
| # ----- Initialization ------ | |
| model_name_or_path = "TheBloke/Llama-2-13B-chat-GGUF" | |
| model_basename = "llama-2-13b-chat.Q5_K_M.gguf" | |
| model_path = hf_hub_download( | |
| repo_id=model_name_or_path, | |
| filename=model_basename | |
| ) | |
| # Initialize the LLAMA model. | |
| llm = Llama( | |
| model_path=model_path, | |
| n_threads=2, # CPU cores | |
| n_batch=512, | |
| n_gpu_layers=43, | |
| n_ctx=4096, | |
| ) | |
| # Load the Whisper model for speech-to-text transcription. | |
| whisper_model = whisper.load_model("base") | |
| # ----- Helper Functions ----- | |
| def transcribe_audio(audio_file): | |
| """Transcribes the provided audio file using Whisper.""" | |
| if audio_file is None: | |
| return "" | |
| result = whisper_model.transcribe(audio_file) | |
| return result["text"] | |
| def generate_response(prompt, max_tokens=150, temperature=0.7): | |
| """ | |
| Uses LLAMA-CPP to generate a response for the given prompt. | |
| """ | |
| output = llm(prompt, max_tokens=max_tokens, temperature=temperature, stop=["User:"]) | |
| response = output["choices"][0]["text"] | |
| return response.strip() | |
| def text_to_speech(text): | |
| """Converts text to speech using gTTS and returns the filepath to the saved audio.""" | |
| tts = gTTS(text=text, lang="en") | |
| tmp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") | |
| tts.save(tmp_file.name) | |
| return tmp_file.name | |
| def voice_chat(audio, text, history, max_tokens, temperature): | |
| """ | |
| Handles a single turn of the conversation: | |
| - If an audio file is provided and no text message, transcribe it. | |
| - Builds a prompt using only the current user input with additional instructions. | |
| - Generates a response from LLAMA. | |
| - Converts the assistant's response to speech. | |
| Returns: | |
| - A new history containing only the current turn. | |
| - The assistant's response text. | |
| - The assistant's response audio filepath. | |
| - Updated state (new history). | |
| """ | |
| # Use the transcribed audio if text is empty. | |
| if audio is not None and (text is None or text.strip() == ""): | |
| user_input = transcribe_audio(audio) | |
| else: | |
| user_input = text if text else "" | |
| # Additional system instructions for improved behavior. | |
| system_prompt = ("You are a helpful, knowledgeable, and concise assistant. " | |
| "Provide accurate, factual, and polite responses. " | |
| "Answer the user's question directly without unnecessary commentary.") | |
| # Build prompt using the system instructions plus the current user input. | |
| prompt = f"{system_prompt}\n\nUser: {user_input}\nAssistant: " | |
| # Generate response using LLAMA-CPP. | |
| response_text = generate_response(prompt, max_tokens=max_tokens, temperature=temperature) | |
| # Convert only the assistant's response to speech. | |
| audio_response = text_to_speech(response_text) | |
| # Reset conversation history to only include the current turn. | |
| new_history = [(user_input, response_text)] | |
| return new_history, response_text, audio_response, new_history | |
| # ----- Gradio Interface ----- | |
| with gr.Blocks() as demo: | |
| gr.Markdown("# Voice Chatbot with LLAMA‑CPP") | |
| with gr.Row(): | |
| with gr.Column(scale=5): | |
| # User inputs: Audio input and/or text input. | |
| audio_input = gr.Audio(type="filepath", label="Speak to Chatbot") | |
| text_input = gr.Textbox(placeholder="Or type your message", label="Your Message") | |
| send_btn = gr.Button("Send") | |
| max_tokens_slider = gr.Slider(50, 300, value=150, step=10, label="Max Tokens") | |
| temperature_slider = gr.Slider(0.1, 1.0, value=0.7, step=0.1, label="Temperature") | |
| with gr.Column(scale=7): | |
| # Display outputs: Chat history, assistant text response, and audio playback. | |
| chat_history = gr.Chatbot(label="Chat History") | |
| response_textbox = gr.Textbox(label="Assistant Response") | |
| audio_output = gr.Audio(label="Response Audio", type="filepath") | |
| # Gradio State to hold the conversation history. | |
| state = gr.State([]) | |
| def run_voice_chat(audio, text, history, max_tokens, temperature): | |
| return voice_chat(audio, text, history, max_tokens, temperature) | |
| # On clicking the "Send" button, run the voice_chat function. | |
| send_btn.click( | |
| fn=run_voice_chat, | |
| inputs=[audio_input, text_input, state, max_tokens_slider, temperature_slider], | |
| outputs=[chat_history, response_textbox, audio_output, state] | |
| ) | |
| # Launch the app. | |
| demo.launch() | |