Meeting transcription now asks which model you want, and has stopped uploading the silence between sentences.
The silence first, because it is pure gain and needs no decision. A recording used to be uploaded with its pauses intact — on a measured recording that was 36 % of everything sent. Those pauses are gone from the upload. Two consequences. Transcription is faster. And a model fed silence tends to invent text to fill it, so a whole class of invented lines disappeared with them. Ours were Russian YouTube subtitle boilerplate: "Продолжение следует…" attributed to somebody who never said it.
Then the choice. Two models, in the Recordings tab under "Transcription model":
- Standard — cheaper, occasionally invents a line.
- Accurate — dearer, noticeably fewer mistakes.
What "fewer mistakes" bought on a test recording: it invented nothing, spelled technical terms correctly, and gave every remark its own timestamp where the cheap one put six sentences under one.
Standard is what everyone gets and it stays the default. Accurate needs a paid plan.
One more thing: Russian speech peppered with English terms comes out better in both models now — "фитчерс флэкс" is written back as "feature flags" instead of being left as the recogniser heard it.