ReadonlyalignmentCross-attention heads (layer, head) used for word timestamps.
ReadonlydimsReadonlydtypeThe checkpoint's weight dtype (mel windows are cast to it).
ReadonlyfastReadonlyistokens [B, T] int32 + audio features → logits [B, T, n_vocab] and, on request, the per-layer
raw cross-attention qk. Advances cache.
Optionalopts: { wantCrossQk?: boolean }
What the transcription engine needs of a speech-recognition graph: the faithful encoder/decoder, the optimized path, and the checkpoint facts window decoding and word timing read. The engine never names the class that implements it.