Who counts speakers on its own, and Topic got three times smaller
Jake Marsh ยท November 21, 2026
Who splits a recording by speaker. It now works out how many people are talking far more reliably, so you can leave num_speakers off: on our meeting test it labels speech with an 18.7% diarization error rate, down from 24.6%, and gets the speaker count about as right as when we tell it the answer. It also uses about a fifth of the CPU time it did, so long recordings come back sooner. Set num_speakers when you know it; it still helps.
Topic tags a web page or document with what it's about and what kind of page it is. It now runs on a model three times smaller and answers about 2.5 times faster. In our tests it names the format (tutorial, news story, product page and so on) correctly 67% of the time instead of 63%, and its topics are as good or better.
Both keep the same options, output and price, so nothing changes in your code. Try them in the playground.