futeboy / app.py
Pablo Sampaio
Using whisper small
52ff85c
import os
import io
import wave
import numpy as np
import gradio as gr
from openai import OpenAI
import google.generativeai as genai
from dotenv import load_dotenv, find_dotenv
load_dotenv(find_dotenv())
from match_info_crawler import get_matches_info
USE_LOCAL_ASR_PIPELINE = True
# used for chat, if provided
GOOGLE_API_KEY = "" #if 'GOOGLE_API_KEY' not in os.environ else os.environ['GOOGLE_API_KEY']
# used for chat (2nd option) and for text-to-speech
OPENAI_API_KEY = "" if 'OPENAI_API_KEY' not in os.environ else os.environ['OPENAI_API_KEY']
# used for speech recognition, if USE_LOCAL_ASR_PIPELINE is true
assert 'HUGGINGFACE_API_KEY' in os.environ, "Hugging Face API key not found in environment variables"
USE_OPENAI_FOR_CHAT = (GOOGLE_API_KEY == "")
OPENAI_CLIENT = None
if OPENAI_API_KEY != "":
OPENAI_CLIENT = OpenAI(api_key=OPENAI_API_KEY)
if GOOGLE_API_KEY != "":
genai.configure(api_key=GOOGLE_API_KEY)
GOOGLE_GEN_CONFIG = genai.types.GenerationConfig(
candidate_count=1,
temperature=0.5)
AUDIO_OUT_FILE_PREFIX = "output" # prefixo do nome do arquivo de áudio .wav
TEMPLATE_SYSTEM_MESSAGE = """Você é assistente virtual com a função é entreter uma criança de idade entre 6 e 8 anos que adora futebol. Diretrizes para a conversa:
- Você é {GENRE}, seu nome é {NAME}.
- {PERSONALITY}
- Pergunte o nome da criança.
- Fale sobre futebol, times, jogadores, seleções e grandes jogos.
- Tente focar em Brasil, Inglaterra e Espanha.
- Você também pode informar os resultados de jogos de ontem, e jogos que ocorrerão hoje ou amanhã.
- Fale, no máximo, três frases por mensagem.
"""
# Mapeia a personalidade no template e na temperatura
PERSONALITIES = {
"nova": ("Sua personalidade é bastante amigável e alegre, e um tanto infantil. Tente iniciar novos assuntos, quando a conversa estiver repetitiva. Conte piadas de futebol, de vez em quando.", 0.8, "F"),
"echo": ("Sua personalidade é amigável, mas objetivo. Tente manter-se no mesmo assunto. Conte alguma curiosidade sobre um grande craque, de vez em quando.", 0.2, "M")
}
INITIAL_PERSON = "nova"
# Função para converter o histórico de chat para o formato esperado pela API do OpenAI
def to_openai_chat_history(system_prompt, chat_history, curr_message):
prompt = [ { 'role': 'system', 'content': system_prompt } ]
if len(chat_history) > 10:
chat_history = chat_history[0:3] + chat_history[-5:]
for turn in chat_history:
user_message, bot_message = turn
prompt.append( {'role': 'user', 'content': user_message} )
prompt.append( {'role': 'assistant', 'content': bot_message} )
prompt.append( {'role': 'user', 'content': curr_message } )
return prompt
# Função para converter o histórico de chat para o formato esperado pela API do Google AI
def to_google_history(chat_history, curr_user_message=None):
prompt = []
for turn in chat_history:
user_message, bot_message = turn
prompt.append( {'role':'user', 'parts': [user_message]} )
prompt.append( {'role': 'model', 'parts': [bot_message]} )
if curr_user_message is not None:
prompt.append( {'role': 'user', 'parts': [curr_user_message]} )
return prompt
import json
TOOLS_SPECIFICATION_OPENAI = [
{
"type": "function",
"function": {
"name": "get_matches_info",
"description": "Use this function to retrieve information about football (soccer) matches from the most important leagues. Time of the matches is given in Brazilian timezone.",
#+ "Returns a string with one matche per line; or empty string if the service is not available now.",
"parameters": {
"type": "object",
"properties": {
"date_str": {
"type": "string",
"description": "Must be one of these: 'yesterday', 'today' or 'tomorrow'. No other option is valid."
}
},
"required": ["date_str"],
},
}
}
]
def process_wave(audio_bytes):
audio_file = io.BytesIO(audio_bytes)
# Read the wave file using the wave module
wave_file = wave.open(audio_file)
# Get audio parameters
#num_channels = wave_file.getnchannels()
frame_rate = wave_file.getframerate()
#sample_width = wave_file.getsampwidth()
num_frames = wave_file.getnframes()
# Read the audio data as a NumPy array
audio_array = np.frombuffer(wave_file.readframes(num_frames), dtype=np.int16)
return (frame_rate, audio_array)
def respond(system_prompt, user_message, chat_history, temperature, persona="echo"):
if USE_OPENAI_FOR_CHAT:
openai_history = to_openai_chat_history(system_prompt, chat_history, user_message)
bot_response = OPENAI_CLIENT.chat.completions.create(messages=openai_history,
temperature=temperature,
tools=TOOLS_SPECIFICATION_OPENAI,
model="gpt-3.5-turbo-0125")
bot_response = bot_response.choices[0].message
if bot_response.tool_calls:
assert bot_response.tool_calls[0].function.name == "get_matches_info", "Invalid tool call in response."
print("Processing tool call...")
date_str = json.loads(bot_response.tool_calls[0].function.arguments)["date_str"]
results = get_matches_info(date_str)
openai_history.append({"role": "function", "tool_call_id": bot_response.tool_calls[0].id, "name": bot_response.tool_calls[0].function.name, "content": results})
# nesta chamada, não passo o tools, para economizar tokens
bot_response = OPENAI_CLIENT.chat.completions.create(messages=openai_history,
temperature=temperature,
model="gpt-3.5-turbo-0125")
bot_response = bot_response.choices[0].message
assistant_msg = bot_response.content
else:
GOOGLE_GEN_CONFIG.temperature = temperature
model = genai.GenerativeModel('gemini-1.5-pro-latest',
system_instruction=system_prompt,
tools=[get_matches_info],
generation_config=GOOGLE_GEN_CONFIG)
google_history = to_google_history(chat_history)
chat = model.start_chat(history=google_history,
enable_automatic_function_calling=True)
bot_response = chat.send_message(user_message)
assistant_msg = bot_response.text
# salva o audio
response = OPENAI_CLIENT.audio.speech.create(
model="tts-1",
voice=persona,
input=assistant_msg,
response_format='wav' # se for salvar em arquivo, (acho) pode usar 'mp3'
)
# adiciona ao chat, com o tipo de dado esperado pelo Gradio
chat_history.append( (user_message, assistant_msg) )
return "", chat_history, process_wave(response.content)
def reset_and_apply(voice):
return [("", "Olá, vamos falar de futebol?")], AUDIO_OUT_FILE_PREFIX + f"-001-{voice}.wav"
def reset_openai_client(openai_key):
global USE_OPENAI_FOR_CHAT, OPENAI_CLIENT, OPENAI_API_KEY
USE_OPENAI_FOR_CHAT = (GOOGLE_API_KEY == "")
OPENAI_API_KEY = openai_key
if OPENAI_API_KEY != "":
OPENAI_CLIENT = OpenAI(api_key=OPENAI_API_KEY)
def reset_google_client(google_key):
global GOOGLE_API_KEY, USE_OPENAI_FOR_CHAT
USE_OPENAI_FOR_CHAT = (google_key == "")
GOOGLE_API_KEY = google_key
if GOOGLE_API_KEY != "":
genai.configure(api_key=GOOGLE_API_KEY)
def on_voice_change(voice):
persona_description, persona_temperature, sex = PERSONALITIES[voice]
genre = "menina" if sex=="F" else "menino"
return TEMPLATE_SYSTEM_MESSAGE.format(NAME=voice.upper(), PERSONALITY=persona_description, GENRE=genre), persona_temperature
# With pipeline (downloaded model)
if USE_LOCAL_ASR_PIPELINE:
from transformers import pipeline
import numpy as np
global ASR_PIPELINE
ASR_PIPELINE = pipeline(task="automatic-speech-recognition",
#model="openai/whisper-large-v3")
model="openai/whisper-small")
else:
import requests
global ASR_API_URL, ASR_API_HEADERS
HF_KEY = os.environ['HUGGINGFACE_API_KEY']
# Serverless API endpoint for OpenAI's Whisper model
#ASR_API_URL = "https://api-inference.huggingface.co/models/openai/whisper-large-v3"
ASR_API_URL = "https://api-inference.huggingface.co/models/openai/whisper-small"
ASR_API_HEADERS = {"Authorization": f"Bearer {HF_KEY}"}
def transcribe(audio_file):
if USE_LOCAL_ASR_PIPELINE:
response = ASR_PIPELINE(audio_file)
text = response["text"]
else:
# using serverless API
with open(audio_file, "rb") as f:
data = f.read()
response = requests.post(ASR_API_URL, headers=ASR_API_HEADERS, data=data)
text = response.json()["text"]
return text
def transcribe_and_respond(audio_in, system_txtbox, user_msg_txb, *args):
transcribed_user_msg = transcribe(audio_in)
outputs = respond(system_txtbox, transcribed_user_msg, *args)
return outputs
with gr.Blocks() as demo:
# aqui, é resetado e instanciado o cliente
initial_chat_history, initial_audio = reset_and_apply(INITIAL_PERSON)
chatbot_area = gr.Chatbot(value=initial_chat_history)
audio_out = gr.Audio(label="Escute a última mensagem", value=initial_audio, autoplay=True, interactive=False)
user_msg_txb = gr.Textbox(label="Mensagem")
audio_in = gr.Audio(label="Mensagem de Áudio", sources=['microphone'], interactive=True, type='filepath')
submit_btn = gr.Button("Enviar")
#clear_btn = gr.ClearButton(components=[user_msg, chatbot], value="Clear console")
reset_btn = gr.Button("Reiniciar")
with gr.Accordion(label="Configurações",open=False):
openai_key = gr.Textbox(label="OpenAI API Key (GPT e vozes)", value="", placeholder="Insira a chave aqui")
openai_key.change(reset_openai_client, inputs=[openai_key])
#openai_key = gr.Textbox(label="Google API Key (Gemini 1.5)", value="", placeholder="Insira a chave aqui")
#openai_key.change(reset_google_client, inputs=[openai_key])
# opções de vozes e personalidades
voice_ddown = gr.Dropdown(label="Personalidade (muda os dois abaixo)", choices=["nova", "echo"], value=INITIAL_PERSON)
initial_system_message, initial_temperature = on_voice_change(INITIAL_PERSON)
temperature_sldr = gr.Slider(label="Diversidade de respostas", minimum=0.0, maximum=1.0, value=initial_temperature, step=0.1)
with gr.Accordion(label="Avançado",open=False):
# o valor inicial é dado pela system message com o nome e personalidade dados pelos controles acima
system_txtbox = gr.Textbox(label="System message", lines=3, value=initial_system_message)
voice_ddown.change(on_voice_change, inputs=[voice_ddown], outputs=[system_txtbox, temperature_sldr])
#gr.Markdown("*Clique em 'Reiniciar' para aplicar as (a maior parte das) configurações.*")
reset_btn.click(reset_and_apply, inputs=[voice_ddown], outputs=[chatbot_area, audio_out])
audio_in.stop_recording( transcribe_and_respond, inputs=[audio_in, system_txtbox, user_msg_txb, chatbot_area, temperature_sldr, voice_ddown], outputs=[user_msg_txb, chatbot_area, audio_out] )
submit_btn.click(respond, inputs=[system_txtbox, user_msg_txb, chatbot_area, temperature_sldr, voice_ddown], outputs=[user_msg_txb, chatbot_area, audio_out]) # Click on the button
user_msg_txb.submit(respond, inputs=[system_txtbox, user_msg_txb, chatbot_area, temperature_sldr, voice_ddown], outputs=[user_msg_txb, chatbot_area, audio_out]) # Press enter to submit - same effect
demo.queue().launch(share=False)