Gregory's Server@bedast Then don't call it AI. Call it speech to text. But if it uses a language model to more effectively predict words based on context rather than doing an analyzable mechanical local transformation, it is at least partly the "bad kind of AI" - it has the capacity to introduce biases from training data making output that "sounds right" but means the wrong thing, which is much worse than substituting nonsensical homophones now and then (which the reader will immediately recognize as mistakes). Same principle as why autocorrected text is worse than text with typos.
|
7 comments
@dalias @bedast speech recognition has used language models for decades now. It was one of original applications of language models, way before they scaled up to aping shakespeare. But even without language models, the act of transcription is very close to generative ai, as its the task of predicting the next text token, given previous tokens and encoded audio sequence.   @dalias @bedast Didn't mathematical/rule-based language modeling start showing massively diminishing returns back like... two~three decades ago or is my information wrong?
As far as I'm aware it would be preferable to start from a rule-based language, and then be able to specifically train a small model on a different captioned sample set of the speaker(s) to eliminate its flakiness. |
@bedast Enthusiastically calling new functionality "AI" signals to your audience that you're aligned with the scams and makes them distrust you.
This is not hard.
If you have privacy respecting, on-device, non-plagiarized, ethically built statistical model based processing, DON'T CALL IT "AI".