app.py

import os
import re
import gradio as gr
import asyncio
import time
import tempfile
from huggingface_hub import InferenceClient
from gtts import gTTS
import speech_recognition as sr

# ...

async def generate_audio(prompt):
    # 음성 인식
    r = sr.Recognizer()
    with sr.Microphone() as source:
        print("Speak:")
        audio = r.listen(source)
        try:
            text = r.recognize_google(audio)
        except:
            return "Could not understand audio"

    # LLM 모델에 입력
    generate_kwargs = dict(
        temperature=0.6,
        max_new_tokens=256,
        top_p=0.95,
        repetition_penalty=1,
        do_sample=True,
        seed=42,
    )
    formatted_prompt = system_instructions1 + text + "[JARVIS]"
    stream = client1.text_generation(
        formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True)
    output = ""
    for response in stream:
        output += response.token.text

    # 음성 출력
    tts = gTTS(output, lang="ko")
    with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp_file:
        tmp_path = tmp_file.name
        tts.save(tmp_path)
    return tmp_path

# ...

with gr.Blocks(css="style.css") as demo:    
    with gr.Row():
        user_input = gr.Textbox(label="Prompt", value="What is Wikipedia")
        input_text = gr.Textbox(label="Input Text", elem_id="important")
        output_audio = gr.Audio(label="JARVIS", type="filepath",
                        interactive=False,
                        autoplay=True,
                        elem_classes="audio")
        translate_btn = gr.Button("Response")
    with gr.Row():
        translate_btn.click(fn=generate1, inputs=user_input,
                            outputs=output_audio, api_name="translate")
        translate_btn.click(fn=generate_audio, inputs=user_input,
                            outputs=output_audio, api_name="generate_audio")
    with gr.Row():
        gr.Markdown(MORE)

if __name__ == "__main__":
    demo.queue(max_size=200).launch()