HuBERT-EE: Early Exiting HuBERT for Efficient Speech Recognition

Self-supervised encoders are deep enough that a large share of utterances finish long before the last layer, and the remaining layers are wasted compute.

Attaches multiple early-exit branches at intermediate layers. Each branch emits a prediction and a confidence score; when the score passes a threshold the forward pass STOPS and that result is returned, so cost becomes data-dependent rather than fixed.

Reports acceleration of HuBERT inference on LibriSpeech while balancing the accuracy/latency tradeoff. The paper does not quote a single speedup factor in the abstract, so no number is recorded here - the mechanism (per-utterance dynamic depth) is the transferable part.

Interspeech 2024 per the arXiv comment. Early exiting is attractive for a batch-1 serving path because it makes mean latency data-dependent, but it is hostile to batched GPU serving: divergent exit depths force padding to the deepest member of the batch, which erases the saving exactly when the GPU is best utilised. Recorded as research-only for that reason - it is a CPU/latency-bound win, not a throughput win. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: interspeech 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.