Factorized Blank Thresholding for Improved Runtime Efficiency of Neural Transducers

TL;DR

This paper introduces a factorized blank thresholding method for RNN-T models that significantly speeds up decoding and reduces power consumption on devices without sacrificing accuracy.

Contribution

It proposes a novel joiner factorization technique that skips expensive computations based on blank probability thresholds, improving runtime efficiency.

Findings

Achieved 26-30% decoding speed-up

Reduced on-device power consumption by 43-53%

No accuracy loss observed

Abstract

We show how factoring the RNN-T's output distribution can significantly reduce the computation cost and power consumption for on-device ASR inference with no loss in accuracy. With the rise in popularity of neural-transducer type models like the RNN-T for on-device ASR, optimizing RNN-T's runtime efficiency is of great interest. While previous work has primarily focused on the optimization of RNN-T's acoustic encoder and predictor, this paper focuses the attention on the joiner. We show that despite being only a small part of RNN-T, the joiner has a large impact on the overall model's runtime efficiency. We propose to utilize HAT-style joiner factorization for the purpose of skipping the more expensive non-blank computation when the blank probability exceeds a certain threshold. Since the blank probability can be computed very efficiently and the RNN-T output is dominated by blanks, our proposed method leads to a 26-30% decoding speed-up and 43-53% reduction in on-device power consumption, all the while incurring no accuracy degradation and being relatively simple to implement.