FastEmit: Low-latency Streaming ASR with Sequence-level Emission Regularization

Streaming ASR emits words late because transducer training optimises full-sequence probability, not emission delay - and existing penalties need external alignments.

Adds a sequence-level emission-latency regulariser directly to the transducer's per-sequence training objective, with no alignment information required. Because it regularises the sequence rather than individual token probabilities, it works across RNN-T, Transformer-T, ConvNet-T and Conformer-T without architecture changes.

150-300 ms latency reduction with significantly better accuracy than prior latency-reduction techniques (Early/Late Penalties, Constrained Alignments) on a Voice Search test set. Baseline methods are named and the test set identified, but hardware is not specified.

ICASSP 2021 per the arXiv comment. Mechanism relevance for inference: emission delay is a DECODING-SCHEDULING property, not a hardware one, so this class of fix is free at serving time once the model is trained with it. Contrast with the local-agreement approach used to retrofit streaming onto Whisper (pap record whisper-streaming): FastEmit bakes latency into training, Whisper-Streaming adds it at inference by re-running the encoder. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: icassp has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.