app.py

import torch
import spaces
import numpy as np
import gradio as gr
from gtts import gTTS
from transformers import pipeline
from huggingface_hub import InferenceClient

# Model names
ASR_MODEL_NAME = "openai/whisper-small"
LLM_MODEL_NAME = "mistralai/Mistral-7B-Instruct-v0.2"

# Initial system prompt
system_prompt = """"<s>[INST] You are Friday, a helpful and conversational AI assistant, and you respond with one to two sentences. [/INST] Hello there! I'm Friday, how can I help you?</s>"""

# Global variables for initial history
initial_instruct_history = system_prompt
initial_formatted_history = ""

# Create inference client for text generation
client = InferenceClient(LLM_MODEL_NAME)

# Set device for ASR pipeline
device = 0 if torch.cuda.is_available() else "cpu"

# ASR pipeline
pipe = pipeline(
    task="automatic-speech-recognition",
    model=ASR_MODEL_NAME,
    device=device,
)

def generate(instruct_history, temperature=0.1, max_new_tokens=128, top_p=0.95, repetition_penalty=1.0):
    temperature = float(temperature)
    if temperature < 1e-2:
        temperature = 1e-2
    top_p = float(top_p)

    generate_kwargs = dict(
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        do_sample=True,
        seed=42,
    )

    output = client.text_generation(
        instruct_history, **generate_kwargs, stream=False, details=False, return_full_text=False)

    return output

@spaces.GPU(duration=60)
def transcribe(audio, instruct_history, formatted_history):
    sr, y = audio
    y = y.astype(np.float32)
    y /= np.max(np.abs(y))

    # Transcribe user audio
    transcribed_user_audio = pipe({"sampling_rate": sr, "raw": y})["text"]

    # Append user input to history
    formatted_history += f"Human: {transcribed_user_audio}\n\n"
    instruct_history += f"<s>[INST] {transcribed_user_audio} [/INST] "

    # Generate LLM response
    llm_response = generate(instruct_history)

    # Append AI response to history
    instruct_history += f" {llm_response}</s>"
    formatted_history += f"Friday: {llm_response}\n\n"

    # Convert AI response to audio
    audio_response = gTTS(llm_response)
    audio_response.save("response.mp3")

    # Return the updated history and audio
    return "response.mp3", formatted_history, instruct_history

with gr.Blocks() as demo:
    gr.HTML("<center><h1>Friday: AI Virtual Assistant</h1><center>")

    # Initialize state
    instruct_state = gr.State(value=initial_instruct_history)
    formatted_state = gr.State(value=initial_formatted_history)

    with gr.Row():
        audio_input = gr.Audio(label="Human", sources="microphone")
        output_audio = gr.Audio(label="Friday", type="filepath", interactive=False, autoplay=True, elem_classes="audio")

    transcribe_btn = gr.Button("Transcribe")

    # Textbox to display the full conversation history
    transcription_box = gr.Textbox(label="Transcription", lines=10, placeholder="Conversation History...")

    # Pass states to the transcribe function and update them after each click
    transcribe_btn.click(
        fn=transcribe,
        inputs=[audio_input, instruct_state, formatted_state],
        outputs=[output_audio, transcription_box, instruct_state, formatted_state]
    )

if __name__ == "__main__":
    demo.queue()
    demo.launch()