- Shipped
- November 6, 2026 at 11:18 PM UTC
- Author
- Kamo
- Commit
- 6fd880a
Owner report 2026-11-06: call transcripts "do not reflect the entirety of the call recordings". Measured on the platform org's calls: the router's first STT choice, gpt-4o-transcribe, returned each <=30 s chunk cut short after a pause (one 29.5 s chunk came back with 8 of its 57 words) and gives no timestamps, so nothing showed where it stopped. On the same six chunks gpt-4o-transcribe-diarize kept 278 words, Whisper 276, gpt-4o-transcribe 185. **************** (with verbose_json) is now accepted: only models TimedTranscription accepts are walked (the Whisper family, diarizing GPT-4o, anything on Speaches), and their segments come back in verbose_json, with a diarizing model's speaker on each. Word timestamps are still refused. A plain request is unchanged.