Whisper feeds its own decoded output forward as the prompt for the next 30s
window, which is what makes it loop once it slips. max_context is the budget for
that carried text; setting it to 0 stops the loops but also disables the initial
prompt, since whisper.cpp gates both on n_max_text_ctx > 0. carry_initial_prompt
puts the prompt in a static slot that is filled first and in full, so making the
prompt longer than the budget leaves nothing over for decoded text — the prompt
reaches every window and none of the model's own output does.
The prompt is a style exemplar, not an instruction: whisper imitates what it is
primed with. A first version described the format ("Commas separate clauses")
and a character surname came back as "Commas", so both prompts are now ordinary
conversational prose with no meta-language, one per language.
Also switch to verbose_json for segment timestamps and rejoin the segments into
paragraphs on pause length, rather than emitting one line per utterance with a
leading space; decode with beam search instead of greedy; widen the VAD segments
so they break at real pauses instead of at breaths; and stop verbose_json from
re-running a full language auto-detect that step 1 already answered.
Measured on a 43 minute episode against the previous output: redundant repeated
sentences 18 -> 9 (the long distinctive loops are gone, what remains is a
catchphrase), paragraphs 1 -> 93, lone-punctuation lines 3 -> 0, lines with a
leading space 1021 -> 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bun's fetch aborts at 300s by default. whisper.cpp runs at roughly 8x realtime,
so a 43-minute recording needs about 5.4 minutes and died with "The operation
timed out" some 20 seconds short of the answer — long enough to look like a slow
machine rather than a ceiling.
AbortSignal.timeout() does not raise that ceiling; only `timeout: false` does,
which the DOM RequestInit type does not declare, hence the local BunRequestInit.
Verified end to end: a 43-minute episode through the task now completes in 547s
with a 27k-character transcript, where it previously failed at exactly 300s.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three changes that Transcribe Audio needs.
Whisper's translate mode only ever outputs English, so it cannot honour
"translate into <language>" for anything else — it answered Portuguese audio with
a rough English rendering instead of a transcript. The translate decision is gone
and transcription is always faithful to the detected language; spokenLanguages
now only breaks ties on clips Whisper is unsure about.
Script tasks get OFFICER_API_URL / OFFICER_API_HOST / OFFICER_AUTH_TOKEN so they
can call Officer's own endpoints rather than reimplementing server-side work.
Requests go to 127.0.0.1 so nothing depends on DNS or the proxy, but origin
validation matches Host against PUBLIC_URL, hence the separate host variable.
`inline` accepts "ask", which offers both affordances in the runner — Run here
streams into the modal, Run as job queues it. Useful when the same task can take
a second or an hour depending on whether it was pointed at a file or a library.
Existing true/false values behave exactly as before.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Language selection ranked only the user's spoken languages, so a user who speaks
one language always got that language. With spoken=["en"], Portuguese audio
detected at 0.968 pt was sent as language=en, and Whisper answered with a rough
English rendering rather than a transcript. `translated` stayed false, so nothing
downstream could tell a translation had happened, and detectedLanguage reported
"en" for audio that was not English.
Take Whisper's top language when it clears 0.5, and fall back to the spoken
languages only when the clip really is ambiguous — which is the case the bias was
written for.
Verified against the local whisper server with pt/en/es recordings:
spoken=["en"] pt -> translated English, es -> translated English, en -> as-is
spoken=["en","pt"] pt -> Portuguese transcript, no translation
spoken=[] pt -> Portuguese transcript
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
pastilhasandClaude Opus 4.8