OptionalaudioOptionalbeamOptionalbestOptionalclip[start, end, start, end, …] seconds; last end defaults to the clip end. Batch only: a run with clips never transcribes eagerly.
OptionalcompressionOptionalconditionOptionalfastExecution path: true (default) runs the optimized graphs
(the graph's fast path) for temperature-0 decoding; false forces the
faithful oracle graph. Sampling (temperature > 0) always uses the
faithful path.
OptionalhallucinationWith wordTimestamps: skip silent gaps longer than this many seconds around a likely hallucination (transcribe.py hallucination_silence_threshold).
OptionalinitialOptionallanguagenull → detect.
OptionallengthOptionallogprobOptionalmaxOptionalnoOptionalobserverParity instrumentation: the decode window's encoder output and each step's PRE-filter last-position logits [n_group, V] f32 (borrowed — valid only during the callback).
OptionalonProgress callback: seeked frames out of content frames.
OptionalonCalled with each finalized segment as the loop emits it.
OptionalpatienceOptionalprefixOptionalpromptOptionalsampleOptionalsignalChecked between windows; an aborted signal throws "transcription cancelled".
OptionalsuppressOptionalsuppressToken ids to suppress; "-1" = the tokenizer's non-speech set (default).
OptionaltaskOptionaltemperatureOptionaltemperaturesSingle temperature or the fallback ladder (oracle default: 0, .2 … 1).
OptionalwithoutOptionalwordWord-level timestamps via cross-attention DTW (packages/inference/src/transcription/whisper/timing.ts); one extra faithful decoder pass per window.
Lab (whisper.cpp
-ac): encode only the firstaudioCtxof the 1500 encoder positions (2·audioCtx mel frames). Cheaper for short clips but NOT the trained context — measured hallucinated repeats below ~1024 on the fixture clips. null = full context. Fast path only.