Modeling accents for automatic speech recognition

Abstract

In British English the term 'accent' refers to systematic variations in pronunciation, often associated with particular geographic regions. Accent is one of the most frequently cited causes of variability in speech. The problem of accents is becoming more important with the advancement of computerised services and Automatic Speech Recognition (ASR) systems. Speech recognition technology is used in a wide range of applications and services such as health care, education, automated call centres, authentication and information services. In many of these applications accents and foreign languages posed a problem for speech recognition developers because, if the people using the system could not be understood, they might become frustrated and stop using the system. For example, a recent news story reported that an automated phone system deployed by Birmingham City Council could not cope with ‘Brummie’ accents [1]. This research investigates how we can exploit the knowledge of accents to obtain both rapid and robust ASR systems for British accented utterances using only 30 seconds of speech.

When we hear another person’s speech for the first time, we quickly establish a 'profile' of that person based on his or her speech, in terms of factors such as gender, age, accent, and social group. It is possible that we use this characterization to adapt very quickly to that person’s speech. Current ASR systems typically differentiate between genders, but otherwise tend to ignore important factors including accents. Additionally, conventional adaptation techniques for ASR require a substantial amount of training material from each individual to be able to adapt the system in order to have a more user specific system. Now the question is whether the concept of ‘regional accent’ is useful to overcome these two problems.

The first section describes the techniques used to visualise the space where accent recognition is performed, otherwise known as the ‘accent recognition’ space, and show the extent to which the emergent structure is consistent with subjective notions of accent. The two techniques used are Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA).

We then interpret speech recognition results on accented speech from 14 different British regions, analysed in terms of their structure in the ‘accent recognition’ space.

Following this we present results on accent adaptation using conventional adaptation techniques such as MAP (Maximum A Posteriori Adaptation) and MLLR (Maximum Likelihood Linear Regression). We use these accent adaptation techniques along with the knowledge from accent recognition techniques, such as the ACCDIST accent recognition measure, to provide an accent robust, rapid speaker adaptation system.

1. “Brummie accents baffle automated phone system ... at Birmingham City Council”, Daily Mail, 5th November 2012.