Quality filters for pretraining data transcripts

[Lauler]

Can you perhaps share details on some of the automated filtering methods you developed to improve transcript quality.

Mostly curious to know in more detail what you actually considered to be "bad transcript quality".

Was a "bad transcription" blatantly mistranscribed or mismatched transcriptions? Or was a "bad transcription" also a transcription that took too many liberties in paraphrasing what was actually said? In the paper you write:

Initial inspection showed a large amount of subpar transcripts in the raw dataset. To address this, we developed several automated filtering methods to improve transcript quality.

I'm curious to know where you set the threshold for subpar transcripts. Does it have to be almost verbatim to pass your quality filter, or can it also be quite heavily paraphrased?