Hell, I just found out a few months ago my father’s office (he’s a lawyer) still did transcriptions of court proceedings manually! Full blown lawyers wasting expensive man hours doing that. As soon as I found out I got some random whisper cpp frontend and slapped it on a random computer
As long as they still have someone verifying and correcting mistakes, that’s a decent solution. It doesn’t need to be complicated. Whisper is by no means perfect, but from my experience it handles 90% fine, only having to listen and correct 10% is a huge improvement over transcribing everything manually.
The issues with so much of the current AI use is implementations that completely remove trained human oversight from the equation because the AI companies position it as a complete replacement for employees.
from my experience it handles 90% fine, only having to listen and correct 10% is a huge improvement over transcribing everything manually.
I’m curious how this is supposed to work? How do you only listen to 10%? How are you supposed to know which 10% to listen to? Wouldn’t you still need to listen to it all to find that 10%?
Yes, but listening to it all once and correcting the mistakes as you hear/see them would be faster than transcribing the entire thing, then proof reading.
Hell, I just found out a few months ago my father’s office (he’s a lawyer) still did transcriptions of court proceedings manually! Full blown lawyers wasting expensive man hours doing that. As soon as I found out I got some random whisper cpp frontend and slapped it on a random computer
I hope his clients are aware
As long as they still have someone verifying and correcting mistakes, that’s a decent solution. It doesn’t need to be complicated. Whisper is by no means perfect, but from my experience it handles 90% fine, only having to listen and correct 10% is a huge improvement over transcribing everything manually.
The issues with so much of the current AI use is implementations that completely remove trained human oversight from the equation because the AI companies position it as a complete replacement for employees.
I’m curious how this is supposed to work? How do you only listen to 10%? How are you supposed to know which 10% to listen to? Wouldn’t you still need to listen to it all to find that 10%?
Yes, but listening to it all once and correcting the mistakes as you hear/see them would be faster than transcribing the entire thing, then proof reading.