Method transcribe

autocut/whisper_model.py:179–268

(
        self,
        input: srt,
        audio: np.ndarray,
        speech_array_indices: List[SPEECH_ARRAY_INDEX],
        lang: LANG,
        prompt: str,
    )

        self.whisper_model = partial(openai.Audio.transcribe, model=model_name)

    def transcribe(
        self,
        input: srt,
        audio: np.ndarray,
        speech_array_indices: List[SPEECH_ARRAY_INDEX],
        lang: LANG,
        prompt: str,
    ) -> List[srt.Subtitle]:
        res = []
        name, _ = os.path.splitext(input)
        raw_audio = AudioSegment.from_file(input)
        ms_bytes = len(raw_audio[:1].raw_data)
        audios: List[
            TypedDict(
                "AudioInfo", {"input": str, "audio": AudioSegment, "start_ms": float}
            )
        ] = []

        i = 0
        for index in speech_array_indices:
            start = int(index["start"]) / self.sample_rate * 1000
            end = int(index["end"]) / self.sample_rate * 1000
            audio_seg = raw_audio[start:end]
            if len(audio_seg.raw_data) < self.split_audio_bytes:
                temp_file = f"{name}_temp_{i}.wav"
                audios.append(
                    {"input": temp_file, "audio": audio_seg, "start_ms": start}
                )
            else:
                logging.info(
                    f"Long audio with a size({len(audio_seg.raw_data)} bytes) greater than 25M({25 * 2 ** 20} bytes) "
                    "will be segmented"
                    "due to Openai's API restrictions on files smaller than 25M"
                )
                split_num = len(audio_seg.raw_data) // self.split_audio_bytes + 1
                for j in range(split_num):
                    temp_file = f"{name}_{i}_temp_{j}.wav"
                    split_audio = audio_seg[
                        j
                        * self.split_audio_bytes
                        // ms_bytes : (j + 1)
                        * self.split_audio_bytes
                        // ms_bytes
                    ]
                    audios.append(
                        {
                            "input": temp_file,
                            "audio": split_audio,
                            "start_ms": start + j * self.split_audio_bytes // ms_bytes,
                        }
                    )
            i += 1

        if len(audios) > 1:
            from multiprocessing import Pool

            pbar = tqdm(total=len(audios))