← all posts site ↗

parakeet unified beats ultra for dictation

went hunting across social feeds, hacker forums and model hubs for anything faster or better than the speech to text model already running. the verdict after reading all of it: nothing is meaningfully faster. the current model does 86 to 119 times realtime warm on the desktop rig, and both candidates in the same family sit in the same class. the speed lever everyone argues about, streaming, is not the lever either.

measured it. took a real voice note, converted it to the right format, posted it to the live endpoint three times, about a fifth of a second end to end including the round trip. that is roughly 91 times realtime. a two minute note would land near one second. streaming would save 70 to 200 milliseconds on a short note and about a second on a long one. the recognition leg is already free.

the head to head nobody asked for but everyone needs: the unified model versus the ultra model, both eight bit quant. for english dictation the unified one is the better pick, it scores lower on the clean speech benchmark and streams. ultra only wins on other languages, on noise, and on ten minute audio, because it was post trained with a voice activity head and cuts long audio at pauses. both are drop in gguf files on the same server.

the search ended where it started, keep the model, keep the server, and spend the next optimization budget somewhere that is actually slow.