It’s improved, but not enough to make it useful as more than an aide for people who can’t type.
I don’t think this is true.
There’s a locally hostable model called whisper that is very impressive.
My plumber uses speech to text to send text messages all day.
Late Night Linux guy says he uses it for microsoft teams quite a bit.
You’re only partially correct about input speed. If you want to dictate an email then yes you need to think about each word you want to say and the order in which to say them. Coupled with an LLM that problem is diminished because you can just kind of have a conversation with the LLM and tell it to draft an email.
You’re only partially correct about input speed. If you want to dictate an email then yes you need to think about each word you want to say and the order in which to say them. Coupled with an LLM that problem is diminished because you can just kind of have a conversation with the LLM and tell it to draft an email.
and how much of that conversation with an LLM is “No, what I want is…” because it assumed something; or just straight up hallucinated or the typo made it go off on a tangent?
As for whisper, I can find sources that are saying for American-English speakers in a not-noisy environment (aka the best case scenario,) the model has a word error rate between 2-8%. For reference, Dragon NaturallySpeaking had a WER of 3-5%. So I wouldn’t say that Whisper has made any substantial improvements, and they’re OpenAi. you can trust them if you want. I don’t think that’ll work out well in the long run, though.
I’d like to see the source that says Dragon’s WER in the 90s was 3-5%. I used Dragon in the 2000s and it just wasn’t comparable to the current state of the art.
whisper.cpp is an opensource implementation, although I’m not certain exactly how open.
when you’re providing context rather than instructions the tendency for a model to hallucinate or run off on a tangent is minimal, because the context you’re providing has it’s own cohesion.
I don’t think this is true.
There’s a locally hostable model called whisper that is very impressive.
My plumber uses speech to text to send text messages all day.
Late Night Linux guy says he uses it for microsoft teams quite a bit.
You’re only partially correct about input speed. If you want to dictate an email then yes you need to think about each word you want to say and the order in which to say them. Coupled with an LLM that problem is diminished because you can just kind of have a conversation with the LLM and tell it to draft an email.
and how much of that conversation with an LLM is “No, what I want is…” because it assumed something; or just straight up hallucinated or the typo made it go off on a tangent?
As for whisper, I can find sources that are saying for American-English speakers in a not-noisy environment (aka the best case scenario,) the model has a word error rate between 2-8%. For reference, Dragon NaturallySpeaking had a WER of 3-5%. So I wouldn’t say that Whisper has made any substantial improvements, and they’re OpenAi. you can trust them if you want. I don’t think that’ll work out well in the long run, though.
I’d like to see the source that says Dragon’s WER in the 90s was 3-5%. I used Dragon in the 2000s and it just wasn’t comparable to the current state of the art.
whisper.cpp is an opensource implementation, although I’m not certain exactly how open.
when you’re providing context rather than instructions the tendency for a model to hallucinate or run off on a tangent is minimal, because the context you’re providing has it’s own cohesion.