app.py

import io
import base64
import threading
import numpy as np
from gtts import gTTS
import streamlit as st
import speech_recognition as sr
from huggingface_hub import InferenceClient
from streamlit_mic_recorder import mic_recorder
import webrtcvad
import wave

temp_audio_file_path = "./output.wav"

if "history" not in st.session_state:
    st.session_state.history = []

if "pre_prompt_sent" not in st.session_state:
    st.session_state.pre_prompt_sent = False

gatherUsageStats = "false"
pre_prompt_text = "eres una IA conductual, tus respuestas serán breves."

def recognize_speech(audio_data, show_messages=True):
    recognizer = sr.Recognizer()

    with io.BytesIO(audio_data) as audio_file:
        try:
            audio_text = recognizer.recognize_google(audio_file, language="es-ES")
            if show_messages:
                st.subheader("Texto Reconocido:")
                st.write(audio_text)
                st.success("Reconocimiento de voz completado.")
        except sr.UnknownValueError:
            st.warning("No se pudo reconocer el audio. ¿Intentaste grabar algo?")
            audio_text = ""
        except sr.RequestError:
            st.error("Hablame para comenzar!")
            audio_text = ""

    return audio_text

def start_stream():
    recognizer = sr.Recognizer()
    with sr.AudioFile(temp_audio_file_path) as source:
        audio = recognizer.record(source)

    audio_data = np.frombuffer(audio.frame_data, dtype=np.int16)
    silent_chunks = detect_silence(audio_data)

    for silent_chunk in silent_chunks:
        audio_text = recognize_speech(silent_chunk)
        st.subheader("Texto Reconocido:")
        st.write(audio_text)

        # Actualiza el historial con el texto reconocido
        st.session_state.history.append((audio_text, ""))

        # Genera la respuesta de texto a voz usando el texto reconocido como prompt
        response, audio_file = generate(audio_text, st.session_state.history)

        # Muestra la respuesta y reproduce el audio
        display_recognition_result(audio_text, response, audio_file)

def detect_silence(audio_data, silence_threshold=5000, silence_duration=5000):
    is_silence = lambda x: max(x) < silence_threshold
    chunks = [audio_data[i:i+silence_duration] for i in range(0, len(audio_data), silence_duration)]
    silent_chunks = [chunk for chunk in chunks if is_silence(chunk)]
    return silent_chunks

def format_prompt(message, history):
    prompt = "<s>"

    if not st.session_state.pre_prompt_sent:
        prompt += f"[INST]{pre_prompt_text}[/INST]"

    for user_prompt, bot_response in history:
        prompt += f"[INST] {user_prompt} [/INST]"
        prompt += f" {bot_response}</s> "

    prompt += f"[INST] {message} [/INST]"
    return prompt

def generate(audio_text, history, temperature=None, max_new_tokens=512, top_p=0.95, repetition_penalty=1.0):
    client = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")

    temperature = float(temperature) if temperature is not None else 0.9
    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,)

    formatted_prompt = format_prompt(audio_text, history)
    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True)
    response = ""

    for response_token in stream:
        response += response_token.token.text
    
    response = ' '.join(response.split()).replace('</s>', '')
    audio_file = text_to_speech(response, speed=1.3)
    return response, audio_file

def text_to_speech(text, speed=1.3):
    tts = gTTS(text=text, lang='es')
    audio_fp = io.BytesIO()
    tts.write_to_fp(audio_fp)
    audio_fp.seek(0)
    return audio_fp

def audio_play(audio_fp):
    st.audio(audio_fp.read(), format="audio/mp3", start_time=0)

def display_recognition_result(audio_text, output, audio_file):
    if audio_text:
        st.session_state.history.append((audio_text, output))  

    if audio_file is not None:
        st.markdown(
            f"""<audio autoplay="autoplay" controls="controls" src="data:audio/mp3;base64,{base64.b64encode(audio_file.read()).decode()}" type="audio/mp3" id="audio_player"></audio>""",
            unsafe_allow_html=True)

def voice_activity_detection(audio_data):
    return vad.is_speech(audio_data, sample_rate)

def start_stream():
    recognizer = sr.Recognizer()
    with sr.AudioFile(temp_audio_file_path) as source:
        audio = recognizer.record(source)

    audio_data = np.frombuffer(audio.frame_data, dtype=np.int16)
    silent_chunks = detect_silence(audio_data)

    for silent_chunk in silent_chunks:
        audio_text = recognize_speech(silent_chunk)
        st.subheader("Texto Reconocido:")
        st.write(audio_text)

audio = mic_recorder(start_prompt="▶️", stop_prompt="🛑", key='recorder')

if audio:
    st.audio(audio['bytes'])

    audio_bytes = audio["bytes"]
    sample_width = audio["sample_width"]
    sample_rate = audio["sample_rate"]

    with wave.open(temp_audio_file_path, 'wb') as wave_file:
        wave_file.setnchannels(1)
        wave_file.setsampwidth(sample_width)
        wave_file.setframerate(sample_rate)
        wave_file.writeframes(audio_bytes)

vad = webrtcvad.Vad(1)
channels = [1]
mapping = [c - 1 for c in channels]
interval_size = 10 
downsample = 1