Using generalized maxout networks and phoneme mapping for low resource ASR- a case study on Flemish-Afrikaans

摘要

Recently, multilingual deep neural networks (DNNs) have been successfully used to improve under-resourced speech recognizers. Common approaches use either a merged universal phoneme set based on the International Phonetic Alphabet (IPA) or a language specific phoneme set to train a multilingual DNN. In this paper, we investigate the effect of both knowledge-based and data-driven phoneme mapping on the multilingual DNN and its application to an under-resourced language. For the data-driven phoneme mapping we propose to use an approximation of Kullback Leibler Divergence (KLD) to generate a confusion matrix and find the best matching phonemes of the target language for each individual phoneme in the donor language. Moreover, we explore the use of recently proposed generalized maxout network in both multilingual and low resource monolingual scenarios. We evaluate the proposed phoneme mappings on a phoneme recognition task with both HMM/GMM and DNN systems with generalized maxout architecture where Flemish and Afrikaans are used as donor and under-resourced target languages respectively.