24
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 20 Jul 2025
24 points (100.0% liked)
TechTakes
2098 readers
241 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 2 years ago
MODERATORS
This incredible banger of a bug against whisper, the OpenAI speech to text engine:
Lol, training data must have included videos where there was silence but on screen was a credit for translation. Silence in audio shouldn't require special "workarounds".
The whisper model has always been pretty crappy at these things: I use a speech to text system as an assistive input method when my RSI gets bad and it has support for whisper (because that supports more languages than the developer could train on their own infrastructure/time) since maybe 2022 or so: every time someone tries to use it, they run into hallucinated inputs in pauses - even with very good silence detection and noise filtering.
This is just not a use case of interest to the people making whisper, imagine that.
Similar case from 2 years ago with Whisper when transcribing German.
I'm confused by this. Didn't we have pretty decent speech-to-text already, before LLMs? It wasn't perfect but at least didn't hallucinate random things into the text? Why the heck was that replaced with this stuff??
Transformers do way better transcription, buuuuuut yeah you gotta check it
I'm just confused because I remember using Dragon Naturally Speaking for Windows 98 in the 90s and it worked pretty accurately already back then for dictation and sometimes it feels as if all of that never happened.
Discovered some commentary from Baldur Bjarnason about this:
On a personal sidenote, I can see non-English text/audio becoming a form of low-background media in and of itself, for two main reasons:
First, LLMs' poor performance in languages other than English will make non-English AI slop easier to identify - and, by extension, easier to avoid
Second, non-English datasets will (likely) contain less AI slop in general than English datasets - between English being widely used across the world, the tech corps behind this bubble being largely American, and LLM userbases being largely English-speaking, chances are AI slop will be primarily generated in English, with non-English AI slop being a relative rarity.
By extension, knowing a second language will become more valuable as well, as it would allow you to access (and translate) low-background sources that your English-only counterparts cannot.
do you keep count/track? the moleskine must be getting full!
I don't keep track, I just put these together when I've got an interesting tangent to go on.