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>
