Avoid Overthinking in Self-Supervised Models for Speech Recognition
This work addresses slow inference and accuracy degradation in speech recognition models, particularly for out-of-distribution data, but is incremental as it builds on existing early exit methods.
The paper tackled the problem of overthinking in self-supervised models for speech recognition, showing that predictions from the last layer are worse than from intermediate layers, and proposed two new early exit strategies that improve performance and speed trade-offs, achieving up to 2.5x speedup with minimal accuracy loss.
Self-supervised learning (SSL) models reshaped our approach to speech, language and vision. However their huge size and the opaque relations between their layers and tasks result in slow inference and network overthinking, where predictions made from the last layer of large models is worse than those made from intermediate layers. Early exit (EE) strategies can solve both issues by dynamically reducing computations at inference time for certain samples. Although popular for classification tasks in vision and language, EE has seen less use for sequence-to-sequence speech recognition (ASR) tasks where outputs from early layers are often degenerate. This challenge is further compounded when speech SSL models are applied on out-of-distribution (OOD) data. This paper first shows that SSL models do overthinking in ASR. We then motivate further research in EE by computing an optimal bound for performance versus speed trade-offs. To approach this bound we propose two new strategies for ASR: (1) we adapt the recently proposed patience strategy to ASR; and (2) we design a new EE strategy specific to ASR that performs better than all strategies previously introduced.