RNN-T For Latency Controlled ASR With Improved Beam Search

TL;DR

This paper explores latency-controlled RNN Transducers for speech recognition, improving beam search decoding speed and demonstrating comparable accuracy with enhanced efficiency on English video datasets.

Contribution

It introduces a latency-tuning mechanism for RNN-T and enhances the beam search algorithm for faster decoding, advancing real-time ASR applications.

Findings

Achieved comparable WER to hybrid systems

Improved decoding speed of RNN-T beam search

Demonstrated efficiency on English video datasets

Abstract

Neural transducer-based systems such as RNN Transducers (RNN-T) for automatic speech recognition (ASR) blend the individual components of a traditional hybrid ASR systems (acoustic model, language model, punctuation model, inverse text normalization) into one single model. This greatly simplifies training and inference and hence makes RNN-T a desirable choice for ASR systems. In this work, we investigate use of RNN-T in applications that require a tune-able latency budget during inference time. We also improved the decoding speed of the originally proposed RNN-T beam search algorithm. We evaluated our proposed system on English videos ASR dataset and show that neural RNN-T models can achieve comparable WER and better computational efficiency compared to a well tuned hybrid ASR baseline.