Thank you! I studied linguistics, this is why I'm aware of this reality. I'm asking this because this is HN and chances that your worked on speech-to-texts syst

Thank you! I studied linguistics, this is why I'm aware of this reality. I'm asking this because this is HN and chances that your worked on speech-to-texts systems is rather high. I was wondering how this problem is tackled in systems such as Whisper. I haven't used it much, but I definitively noticed that YouTube's transcription model try to strike a balance between transcription fidelity and readability. Sometimes it will elude some pauses & repairs to produce a more streamlined transcription, sometimes it will include them. It's an art I guess, but a model where you could tune this at inference time would be interesting.