I’m Deaf and I use cochlear implants. Captioning is not a side feature for me. It is often how I get the information everyone else already heard.

That makes a dropped name, time, dosage, address, or warning different from an ordinary typo. The rest of a transcript can be perfect and the caption can still fail at the exact moment I need it.

A transcript is useful when it preserves what matters, not only when most of its words match.

The gap I kept seeing

Word error rate is useful, but it treats every word as part of one average. It does not know that “15 milligrams, not 50” is more important than a missing filler word. It also does not tell us whether an alarm, knock, or ringing phone was captioned at all.

I made DeafBench to test those failures directly. It measures word errors, but it also checks critical information, environmental sounds, speaker attribution, and latency.

What I want from it

I want people building speech systems to see where a clean score hides an accessibility problem. I also want the work to be reproducible, so a claim points back to the samples, model revision, predictions, and report that produced it.

DeafBench is open source. Read the code, run the examples, and question the results on GitHub.

This is still early work. I’m learning, testing, and correcting it as I go. That is part of the point: make the evidence visible enough that other people can help improve it.