Incorrect timing

[bullhorn-admin]

Hello, in the following example I say "Purple submarine" between 7.4s-8.6s

0bcad4b4-04e3-4794-907e-1aa59b3ff9b2-2.mp3.zip

However openai whisper thinks that I stopped at 7.0

{
    "text": "''Purple Submarine'",
    "task": "transcribe",
    "language": "english",
    "duration": 20.0,
    "segments": [
        {
            "id": 0,
            "seek": 0,
            "start": 0.0,
            "end": 7.0,
            "text": " ''Purple Submarine'",
            "tokens": [
                50364,
                12119,
                47,
                374,
                781,
                8511,
                6209,
                533,
                6,
                50714
            ],
            "temperature": 0.0,
            "avg_logprob": -0.8815329074859619,
            "compression_ratio": 0.7037037014961243,
            "no_speech_prob": 0.14761993288993835
        }
    ]
}

This is a big deal for me as I am unable to attribute it to the correct speaker.

Any ideas what it is all about?

Thank you,
Eugene

[ryanheise]

Settings the word_timestamps option to True tends to give more accurate timestamps. e.g. From the command line, add the option --word_timestamps True

[misutoneko]

Always be using --suppress_tokens

Well, that's what I do anyway. Small model with --suppress_tokens 50364 seems to get it right in this particular case.

[bullhorn-admin]

I am using OpenAI version in their cloud. I assume there is no way to try what you suggested there, so I need to switch to self hosted version?