Controllable Verbatim ASR via Transcription Policy Latent Variables
July 20, 2026
A new method uses coverage-aware decoder task tokens to control transcription style, addressing decoding instability caused by verbatim versus intended speech mismatch. This approach raised German disfluency F1 from 10% to 79% zero-shot using English-only training.
HOW THIS AFFECTS YOU
●
builderYou can implement more reliable word-level timing and controllable verbatim transcription in ASR pipelines.
●
researcherYou can utilize task tokens to mitigate evaluation confounding caused by transcription style mismatch.