Effect of Neural Network based phonetic feature segmentation in ASR

Abstract

This paper describes a system for phone segmentation using phonetic features, where context information influences the performance of Automatic Speech Recognition (ASR). Current Hidden Markov Model (HMM) based ASR systems have solved this problem by using context-sensitive triphone models. However, these models need a large number of speech parameters and a large volume of speech corpus. In this paper, we propose a technique to model a dynamic process of co-articulation and embed it to ASR systems. Recurrent Neural Network (RNN) is expected to realize this dynamic process. But main problem is the slowness of RNN for training the network of large size. We introduce Distinctive Phonetic Feature (DPF) based feature extraction using a two-stage system consists of a Multi-Layer Neural Network (MLN) in the first stage and another MLN with longer context window in the second stage where the first MLN is expected to reduce the dynamics of acoustic feature pattern and the second MLN to suppress the fluctuation caused by DPF context. The experiments are carried out using Japanese triphthong and Japanese Newspaper Article Sentences (JNAS) data. The proposed DPF based feature extractor provides better segmentation performance with a reduced mixture-set of HMMs. Better context effect is achieved with less computation using MLN instead of RNN.