My agent used 15,000 of my old dictations to fix the words my speech tool kept getting wrong
He turned his 15,000 saved Wispr Flow dictations into a test set for the local Whisper model behind his AI Agent. The agent made find and fix rules from the words Whisper kept mishearing, tested each rule against the full history, and got hard words right far more often: 99 of 240 up to 183 of 240 on old clips, and 19 of 45 up to 43 of 45 on sentences he read out loud.
How it works
- Take 15,000 saved dictations as a test set
- Turn his Wispr dictionary into a hint list
- Compare raw and corrected text to find words Whisper keeps mishearing
- Make find and fix rules and test every rule against the full history
- Hard words go from 99 of 240 right to 183 of 240, and read aloud sentences from 19 of 45 to 43 of 45
- Agent now saves the data every night and looks for new fixes every week
Why it’s worth knowing
It never retrained the model. It learned from 15,000 of his own past mistakes, tested each fix against all of them, and nearly doubled the score.
Where it’s from
Summarized from a public post on whatagentscando.com. See the original ↗