Big model only for hard audios: Sample dependent Whisper model selection for efficient inferences

TL;DR

This paper introduces a sample-dependent model selection method for ASR that chooses the smallest sufficient Whisper model for each audio sample, enabling efficient inference with minimal performance loss.

Contribution

It proposes a decision module that dynamically selects the appropriate model size per sample, reducing computational cost in ASR systems.

Findings

Significant computational savings achieved

Minimal performance degradation with sample-dependent selection

Effective application to Whisper models of different sizes

Abstract

Recent progress in Automatic Speech Recognition (ASR) has been coupled with a substantial increase in the model sizes, which may now contain billions of parameters, leading to slow inferences even with adapted hardware. In this context, several ASR models exist in various sizes, with different inference costs leading to different performance levels. Based on the observation that smaller models perform optimally on large parts of testing corpora, we propose to train a decision module, that would allow, given an audio sample, to use the smallest sufficient model leading to a good transcription. We apply our approach to two Whisper models with different sizes. By keeping the decision process computationally efficient, we build a decision module that allows substantial computational savings with reduced performance drops.