Lexic
A speech recognition engine that can be taught the words your work actually uses. This page is a demonstration of it — talk, and it writes you a note. The interesting part is what happens to the words in between.
What is hard about this
It hears jargon correctly
Recognisers fail on specialist vocabulary in a particular way: they return something that sounds right and is spelled wrong. Pick a field above and Lexic repairs those words by sound, against that field's vocabulary.
It knows when it is unsure
Every word carries the engine's own confidence in what it heard. Lexic only overrules a word it heard poorly — so it can hold a large vocabulary without ever rewriting speech it got right the first time.
It shows its work
Every correction appears beneath the note: what was heard, what replaced it, and how close the match was. Nothing is changed quietly behind your back.
It runs on ordinary hardware
The listening happens on this server's processor. No graphics card, no transcription service, no per-minute fee — which is what makes it practical to run somewhere you control.
Your voice is not kept
The audio is processed by the engine as it arrives and discarded. What remains is the text — the words, not the recording of you saying them.
It never loses what you said
If anything downstream is unavailable, your words are still kept. A missing summary is an inconvenience; a lost recording is not.
Where your words go
| Listening | — |
| Writing the note | — |
| In short | — |
| Your audio | — |
| Your transcript | — |
| Longest recording | — |