Speculative ASR drafts lose accuracy between verification steps because their audio attention drifts from the correct position, and runtime or training-time repositioning recovers part of the lost speed.
SpecASR: Accelerating LLM-based Automatic Speech Recognition via Speculative Decoding , year=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.SD 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Alignment Drift in Single-Model Speculative Decoding for ASR: Mechanism, Correction, and Cost
Speculative ASR drafts lose accuracy between verification steps because their audio attention drifts from the correct position, and runtime or training-time repositioning recovers part of the lost speed.