raspberry pi 5 with llamafile
현재 Raspberry pi 5 로 llamafile 을 이용하여 query-answer 를 완료 하였다.
아래 코드에 대해 조금 더 자세히 분석, 해석, 동작 원리를 파악한 후,
언어, 소리 등을 계속 업그레이드 진행 할 예정.
- llamafile
wget https://huggingface.co/jartine/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/TinyLlama-1.1B-Chat-v1.0.Q5_K_M.llamafile?download=true
result
-
Query
Transcription: I am little bit sad.
https://drive.google.com/file/d/1Bjzp709fKFSBplegLYPFHxS_iy7cb3tn/view?usp=drive_link -
Answer
LLM result: ChatCompletionMessage(content="AI assistant: I understand your emotions. It's normal to feel sad sometimes. However, it's essential to remember that sadness is a natural part of life. It's okay to feel sad, and it's also okay to seek help when you need it. Remember that you can always reach out to your AI assistant for support and guidance.", role='assistant', function_call=None, tool_calls=None)
https://drive.google.com/file/d/1BrSXFatmoFMo1TGeQM4jX2mlh1gUaKhV/view?usp=drive_link
code
import time
import math
from io import BytesIO
import pyaudio
import wave
import RPi.GPIO as GPIO
from gtts import gTTS
from pygame import mixer
from openai import OpenAI
import whisper
BUTTON = 8
GPIO.setmode(GPIO.BOARD)
GPIO.setup(BUTTON, GPIO.IN, pull_up_down=GPIO.PUD_UP)
SPEECH_RECOG_MODEL = whisper.load_model("base")
OPENAI_CLIENT = OpenAI(
base_url="http://127.0.0.1:8080/v1",
api_key = "sk-no-key-required"
)
'''
0 {'index': 0,
'structVersion': 2,
'name': 'USB PnP Audio Device: Audio (hw:2,0)',
'hostApi': 0,
'maxInputChannels': 1,
'maxOutputChannels': 2,
'defaultLowInputLatency': 0.007979166666666667,
'defaultLowOutputLatency': 0.007979166666666667,
'defaultHighInputLatency': 0.032,
'defaultHighOutputLatency': 0.032,
'defaultSampleRate': 48000.0}
'''
CHUNK = 4096
FORMAT = pyaudio.paInt16
CHANNELS = 1
# In digital audio, 44,100 Hz (alternately represented as 44.1 kHz) is a common sampling frequency.
RATE = 44100
RECORD_SECS = 5
def record_wav(query_voice_file):
# show details of devices.
audio = pyaudio.PyAudio()
# create pyaudio object.
stream = audio.open(format = FORMAT,
rate = RATE,
channels = 1,
input = True,
frames_per_buffer=CHUNK)
print("start to record voice")
frames = []
# loop through stream and append audio chunks to frame array.
print(int( (RATE / CHUNK) * RECORD_SECS))
for index_rec in range(0, int( (RATE / CHUNK) * RECORD_SECS)):
data = stream.read(CHUNK)
frames.append(data)
print("finished recording")
# stop the stream, close it, and terminate the pyaudio instantiation.
stream.stop_stream()
stream.close()
audio.terminate()
# save the audio frames as .wav file.
wavefile = wave.open(query_voice_file,'wb')
wavefile.setnchannels(CHANNELS)
wavefile.setsampwidth(audio.get_sample_size(FORMAT))
wavefile.setframerate(RATE)
wavefile.writeframes(b''.join(frames))
wavefile.close()
return
def text_to_speech(answer_text, answer_sound_file):
# convert text to sound bytes.
sound_byte = BytesIO()
tts = gTTS(text=answer_text, lang="en-uk")
tts.write_to_fp(sound_byte)
# save sound bytes.
tts.save(answer_sound_file)
# calculate the sound length from save file.
mixer.init()
sound_info = mixer.Sound(answer_sound_file)
sleep_time = math.ceil(sound_info.get_length())
sound_byte.seek(0)
# play sound during the length of the voice.
mixer.music.load(sound_byte, "mp3")
mixer.music.play()
time.sleep(sleep_time)
def main():
# record voice
query_voice_file = 'query.wav'
record_wav(query_voice_file=query_voice_file)
# convert voice to text with openai whisper
result = SPEECH_RECOG_MODEL.transcribe(query_voice_file, fp16=False)
print("Transcription: {0}".format(result["text"]))
request = result["text"]
# query llamafile (LLaMA_CPP)
completion = OPENAI_CLIENT.chat.completions.create(
model="LLaMA_CPP",
messages=[
{"role": "system", "content": "You are an AI assistant. Your priority is helping users with their requests."},
{"role": "user", "content": request}
]
)
# convert text to speech with espeak
print("LLM result: {0}".format(completion.choices[0].message))
# use gTTS instead of espeak.
answer_sound_file = "answer.mp3"
answer_text = completion.choices[0].message.content
text_to_speech(answer_text, answer_sound_file)
# os.system('espeak "{0}" 2>/dev/null'.format(completion.choices[0].message.content))
if __name__ == "__main__":
print("Ready...")
while(True):
if GPIO.input(BUTTON) == GPIO.LOW:
print("Button pressed")
main()
