Amortized Neural Networks for Low-Latency Speech Recognition

TL;DR

This paper introduces Amortized Neural Networks (AmNets) for low-latency speech recognition, enabling dynamic switching between encoder branches to significantly reduce inference cost and latency while maintaining accuracy.

Contribution

The paper proposes AmNets with a novel dynamic switching mechanism for RNN-T models, reducing compute cost and latency in speech recognition tasks.

Findings

Reduced inference cost by up to 45%

Achieved near real-time latency

Maintained accuracy with variable compute architectures

Abstract

We introduce Amortized Neural Networks (AmNets), a compute cost- and latency-aware network architecture particularly well-suited for sequence modeling tasks. We apply AmNets to the Recurrent Neural Network Transducer (RNN-T) to reduce compute cost and latency for an automatic speech recognition (ASR) task. The AmNets RNN-T architecture enables the network to dynamically switch between encoder branches on a frame-by-frame basis. Branches are constructed with variable levels of compute cost and model capacity. Here, we achieve variable compute for two well-known candidate techniques: one using sparse pruning and the other using matrix factorization. Frame-by-frame switching is determined by an arbitrator network that requires negligible compute overhead. We present results using both architectures on LibriSpeech data and show that our proposed architecture can reduce inference cost by up to 45\% and latency to nearly real-time without incurring a loss in accuracy.