Whisper transcription included a random watermark

[rdanielo]

I am transcribing some audios that I have from my family.
They are on Spanish, and whisper is generally transcribing them well without any additional configuration, just send the audio and get a “mostly correct” transcription back.

However, from time to time the output includes some weird artifact’s.
Recently, after a correct transcription, at the end of the text there was an additional note that looked like a subtitles watermark. Something like “Spanish subtitles provided by Xxx.xx community”

It is not a big problem, but it left me wondering how is this happening and why.

[_j]

Weird whisper transcription links to FEMA.gov

Someone else on Reddit yesterday got a similar oddball at the end of their whisper when talking to ChatGPT, something like “this transcription service provided by xxxx company”. I suspect if you train an AI on labeled audio data, and there is enough reinforcement that after transcripts seem finished, metadata is also included and rewarded, you get that metadata showing who’s efforts you stole.

'Transcription Outsourcing, LLC' repeated throughout whisper transcript

Unsanitized data being used as the reward-training of audio-to-text modalities results in activation of that training inference on similar inputs. The highest occurrence of replaying such ancillary metadata would be from under 30-second clips that appear to wrap up, reaching a conclusion, and mirror the types of data that Ditto Transcripts would perform on public data that could be scraped by an AI company thirsty for labeled audio data, such as government works for the hearing-impaired.