whisper return_timestamps="word" giving quite irregular timestamps

[alanwilter]

The timestamps at rows 34 and 35 are completely irregular.

Is it a bug? Is there something I can about it?

Here's how I'm running:

   def transcribe_by_whisper(self, audio_path: str) -> None:
        """Transcribe the audio file by OpenAI whisper (tiny)

        Args:
            audio_path (str): The filepath of the audio clip.

        Notes:
            - The transcription result is stored in Face.transcription_result.
        """
        device = "cuda:0" if torch.cuda.is_available() else "mps:0" if torch.backends.mps.is_available() else "cpu"
        pipeline = transformers.pipeline(
            task="automatic-speech-recognition",
            model="openai/whisper-tiny.en",
            chunk_length_s=10,
            stride_length_s=2,
            device=device,
        )
        self.transcription_result = pipeline(audio_path, return_timestamps="word")