whisper_loop.py

import struct
import threading
import time
import wave

import numpy as np
import pyaudio
from openai import OpenAI

CHUNK = 1000
FORMAT = pyaudio.paInt16
CHANNELS = 1
RATE = 48000
RMSE_THRESHOLD = 0.01


class RecordThread(threading.Thread):
    """収録音声の録音用のスレッド"""

    def __init__(self, filename, verbose=False):
        super(RecordThread, self).__init__()
        self.filename = filename
        self.stop_event = threading.Event()
        self.daemon = True
        self.verbose = verbose
        self.buffer = np.array([])

    def run(self):
        p = pyaudio.PyAudio()

        if self.filename is not None:
            wf = wave.open(self.filename, "wb")
            wf.setnchannels(CHANNELS)
            wf.setsampwidth(p.get_sample_size(FORMAT))
            wf.setframerate(RATE)

        stream = p.open(
            format=FORMAT,
            channels=CHANNELS,
            rate=RATE,
            input=True,
            frames_per_buffer=CHUNK,
        )

        while not self.stop_event.is_set():
            frames = stream.read(CHUNK)

            if self.filename is not None:
                wf.writeframes(b"".join([frames]))

            frames = bytes2array(frames)
            if self.filename is None:
                self.buffer = np.concatenate((self.buffer, frames), axis=None)

        # 音声ファイルを保存
        stream.stop_stream()
        stream.close()
        p.terminate()
        wf.close()

    def stop(self):
        self.stop_event.set()


def bytes2array(frames):
    count = len(frames) / 2
    format = "%dh" % count
    frames = struct.unpack(format, frames)
    frames = np.array([sample / 32768.0 for sample in frames])
    return frames


def main():
    client = OpenAI()

    record_thread = None

    while True:
        record_thread = RecordThread("temp.wav")
        record_thread.start()

        if record_thread is not None and record_thread.is_alive():
            # 入力があるまで録音したまま待機
            input()
            record_thread.stop()

            # スレッドが音声をファイルに書き込むまで待機
            time.sleep(0.5)

            # 音声ファイルに対する処理をここに入れる
            transcript = client.audio.transcriptions.create(
                model="whisper-1", file=open("temp.wav", "rb"), response_format="text"
            ).strip()

            print(transcript)


if __name__ == "__main__":
    main()