Modular neural networks exploit large acoustic context through broad-class posteriors for continuous speech recognition

Abstract

Traditionally, neural networks such as multi-layer perceptrons handle acoustic context by increasing the dimensionality of the observation vector, in order to include information of the neighbouring acoustic vectors, on either side of the current frame. As a result the monolithic network is trained on a high multi-dimensional space. The trend is to use the same fixed-size observation vector across the one network that estimates the posterior probabilities for all phones, simultaneously. We propose a decomposition of the network into modular components, where each component estimates a phone posterior. The size of the observation vector we use, is not fixed across the modularised networks, but rather accounts for the phone that each network is trained to classify. For each observation vector, we estimate very large acoustic context through broad-class posteriors. The use of the broad-class posteriors along with the phone posteriors greatly enhance acoustic modelling. We report significant improvements in phone classification and word recognition on the TIMIT corpus. Our results are also better than the best context-dependent system in the literature.

Pronunciation analysis for children with speech sound disorders

“…Existing speech-analysis technology uses, almost exclusively, phoneme-probability scores that are output by a conventional speech recognizer. Given state-of-the-art automatic phoneme recognition accuracy of 76% on speech from non-hearing-impaired adults [4] and increased acoustic variability observed in children's speech [5], it is not surprising that the success of this phoneme-recognition approach has been limited.…”

Comparative Study of different types of RNN in Speech Classification

“…Results of experiments have shown that the specific deep neural network outperformed the single DNN based speech enhancement with an accuracy of 94.1%. Christos Antoniou in [10] has proposed a new design for a broad classification by the modular neural network where the observation vector was not fixed in size. Phones have been divided into seven classes (vowels, plosives, fricatives, nasals, diphthongs, semi-vowels, closures).…”

Improved clustered hierarchical tandem system with bottom-up processing

“…Different features such as HATS [6], TRAPS [5] and MRASTA [7] carrying temporal information were shown to be complementary to short term features. The hierarchical or parallel structure MLPs [8,10,11,12,13,14] and MLPs with two or three hidden layers [9,15] were also shown to achieve better performance.…”