DuRep: Dual-Mode Speech Representation Learning via ASR-Aware Distillation

TL;DR

DuRep introduces a dual-mode speech encoder that efficiently handles both streaming and non-streaming tasks with state-of-the-art performance, without extra parameters or adjustments.

Contribution

The paper presents DuRep, a novel dual-mode speech encoder capable of operating in both offline and online modes simultaneously, improving performance across speech tasks.

Findings

DuRep-200M improves streaming and non-streaming ASR by over 11%.

Scaling to 2B parameters sets new benchmarks in ASR and related tasks.

Analysis shows trade-offs between acoustic and semantic info across layers.

Abstract

Recent advancements in speech encoders have drawn attention due to their integration with Large Language Models for various speech tasks. While most research has focused on either causal or full-context speech encoders, there's limited exploration to effectively handle both streaming and non-streaming applications, while achieving state-of-the-art performance. We introduce DuRep, a Dual-mode Speech Representation learning setup, which enables a single speech encoder to function efficiently in both offline and online modes without additional parameters or mode-specific adjustments, across downstream tasks. DuRep-200M, our 200M parameter dual-mode encoder, achieves 12% and 11.6% improvements in streaming and non-streaming modes, over baseline encoders on Multilingual ASR. Scaling this approach to 2B parameters, DuRep-2B sets new performance benchmarks across ASR and non-ASR tasks. Our analysis reveals interesting trade-offs between acoustic and semantic information across encoder layers.