Repository navigation

feat(server): honour response_format in /v1/audio/transcriptions (stacked on #794) - #811

Closed

tawei-amd wants to merge 10 commits into

mainfrom

fix/whisper-response-format

Closed

feat(server): honour response_format in /v1/audio/transcriptions (stacked on #794)#811
tawei-amd wants to merge 10 commits into
mainfrom
fix/whisper-response-format

Commits on Oct 17, 2026
- fix(whisper): stop forcing a timestamp every 16 text tokens
  
  quine00
  authored and
  tawei-amd
  committed
  d98ead3
- fix(whisper): feed the language token to the decoder
  
  quine00
  authored and
  tawei-amd
  committed
  0b0c509
- feat(server): honour language and prompt in /v1/audio/transcriptions
  
  quine00
  authored and
  tawei-amd
  committed
  2dcae2c
- fix(whisper): apply Whisper's timestamp rules, <|30.00|> included
  
  quine00
  authored and
  tawei-amd
  committed
  20c0173
- fix(whisper): do not decode the tail of a full window twice
  
  quine00
  authored and
  tawei-amd
  committed
  c713903
- fix(server): do not log transcripts by default
  
  quine00
  authored and
  tawei-amd
  committed
  19885ae
- fix(whisper): drop a held segment's opening timestamp with its text
  
  tawei-amd
  and
  claude
  committed
  6350e88
- fix(whisper): update the Whisper engine and kernels for prompt accuracy
  
  tawei-amd
  and
  claude
  committed
  f4f9341
Commits on Oct 18, 2026
- feat(server): honour response_format in /v1/audio/transcriptions
  
  tawei-amd
  and
  claude
  committed
  cdc8299
- ci(asr): require verbose_json and timestamp_rules
  
  tawei-amd
  and
  claude
  committed
  871ddcf