On improving discrimination capability of an RNN based recognizer

Abstract

This paper presents a set of eective and ecient techniques to improve the discrimination capability of a recurrent neural network (RNN) based isolated word recognizer. The recognizer contains a set of individually trained RNN speech models (RSMs). Each of them represents a dierent word in the vocabulary. Speech recognition is performed by selecting the RSM that best matches the input utterance. For temporal supervised training of the RSMs, a new error function is introduced, in which the contributions of all phonetic components are equalized regardless of their dierence in duration. The learning rate for recurrent connections is amplied. This is aimed at strengthening temporal dependency in the RSMs to capture dynamic characteristics of speech signals. Furthermore, a hierarchical training strategy is employed to facilitate more efcient discriminative training among the RSMs. A series of speaker-dependent recognition experiments are performed to evaluate the eectiveness of the proposed techniques.

Cantonese syllable recognition using neural networks

“…During the first stage, each RSM is trained independently with positive examples, i.e., utterances carrying the designated syllable. Temporal supervised training is carried out using the real-time-recurrent-learning (RTRL) algorithm proposed by Williams and Zipser [23] and the following error function is minimized [24]:…”

“…By using the proposed discriminative training algorithm, the accuracy was improved greatly to 87.1%. On the other hand, the RSM based recognition algorithm was applied to the recognition of 20 English words (ten isolated digits plus ten command words) and an accuracy of 91.9% was achieved in the multispeaker case [12], [24].…”