Early Attentive Sparsification Accelerates Neural Speech Transcription

TL;DR

This paper introduces an early signal sparsification method in transformer-based speech models that significantly accelerates transcription without accuracy loss by exploiting the interpretability of self-attention.

Contribution

It systematically explores sparsification at different encoder layers and compression ratios, finding optimal early-stage sparsification strategies for faster speech transcription.

Findings

Achieves up to 1.6x speedup in English speech transcription.

Optimal sparsification occurs at 40-60% sparsity early in encoding.

No fine-tuning needed for acceleration.

Abstract

Transformer-based neural speech processing has achieved state-of-the-art performance. Since speech audio signals are known to be highly compressible, here we seek to accelerate neural speech transcription by time-domain signal sparsification early in the neural encoding stage, taking advantage of the interpretability of the self-attention mechanism in transformer audio encoders. With the Whisper family of models, we perform a systematic architecture search over the joint space of sparsification stage (a certain encoder layer) and compression ratio (sparsity). We found that the best resulting solutions under 1% accuracy degradation choose to sparsify the hidden state to 40-60% sparsity at an early encoding stage, and thereby achieve up to 1.6x runtime acceleration in English speech transcription tasks on Nvidia GPUs without any fine-tuning.