Speculative Speech Recognition by Audio-Prefixed Low-Rank Adaptation of Language Models

low-latency speech recognition, speculative speech recognition, prefix language model, low-rank adaptation

This paper explores speculative speech recognition (SSR), where we empower conventional automatic speech recognition (ASR) with speculation capabilities, allowing the recognizer to run ahead of audio. We introduce a metric for measuring SSR performance and we propose a model which does SSR by com bining a RNN-Transducer-based ASR system with an audioprefixed language model (LM). The ASR system transcribes ongoing audio and feeds the resulting transcripts, along with an audiodependent prefix, to the LM, which speculates likely completions for the transcriptions. We experiment with a variety of ASR datasets on which show the efficacy our method and the feasibility of SSR as a method of reducing ASR latency.