The part I care about isn’t the word error rate, it’s the disfluency cleanup. I dictate notes on a boat with the engine running and half of what comes back is “um, let’s meet Tuesday—no, Wednesday.” A model that resolves the self-correction and drops the ums is doing the editing pass I currently do by hand.
Two things I’d actually test. Custom vocabulary, because no general model has ever spelled a chartplotter brand right. And the Live API with sub-second latency, which is the difference between voice as a toy and voice as an input method. It’s public preview, so I’d wire it behind an interface I can swap out — a Cloudflare Worker in front, my own fallback behind it. Never let a preview endpoint become load-bearing.
The story — Google introduced Gemini 3.5 Transcribe, a speech-to-text model that converts raw audio into formatted text, handling background noise, jargon, self-corrections and filler words. It ships as two APIs: real-time streaming with sub-second latency, and pre-recorded processing with speaker attribution and word-level timestamps. Artificial Analysis measured 4.0% WER streaming and 2.6% non-streaming, with 85+ languages. Public preview via Google AI Studio. (Source)