Beating Henry Higgins at His Own Game: A Markovian Approach to Dialectology

1. Introduction The performance of speech recognition algorithms degrades considerably due to speaker variability. Aside from gender, the largest cause for speaker variability is accent. If the accent of a speaker can be determined automatically, then accent-specific speech recognition models can be used, thereby increasing speech recognition accuracy. In this study, the problem of accent classification from a database of Central New York (CNY), North India (IND), and Singapore (SIN)-native English speakers is considered. Linguists have studied regional accents extensively and have described many features particular to the three regions we are considering [1]. The problem of automatic accent classification has received attention within the last decade. Vonwiller, Blackburn, and King used artificial neural networks (ANN) for automatic accent classification [2]. Hansen and Arslan found that energy, duration, and spectral information are good features for accent detection, and that the most distinct features of accent are at the phonemic level. They also formulated a Hidden Markov Model (HMM) classification algorithm [3]. Arslan and Hansen further investigated the HMM for accent classification, using three different scenarios: isolated word – full search, continuous speech – full search, and continuous speech – partial search. In all scenarios, a left-to-right HMM topology with no skip states was used. In tests using speech samples from native speakers of American English, Turkish, Chinese, and German, the HMM algorithm was found to classify better than human classifiers [4]. Teixeira, Trancoso, and Serralheiro applied a parallel set or ergodic nets with context independent HMM units to the problem of English accent identification of speakers from 6 different European countries [5]-[6]. Kat and Fung used phoneme-class based HMMs, rather than phoneme-based HMMs for classification [7]. Chen et al. use a Gaussian mixture model (GMM) to classify different accents in Chinese [8]. We have decided to use two different learning systems to perform the task of accent classification. The first is a HMM with a left-to-right phoneme-based topology. The second is a k-nearest neighbor (KNN) algorithm. We will compare the performance of these two learning systems with the performance of human classifiers.