Acoustic approaches to gender and accent identification

Abstract

There has been considerable research on the problems of speaker and language recognition from samples of speech. A less researched problem is that of accent recognition. Although this is a similar problem to language identification, different accents of a language exhibit more fine-grained differences between classes than languages. This presents a tougher problem for traditional classification techniques. In this thesis, we propose and evaluate a number of techniques for gender and accent classification. These techniques are novel modifications and extensions to state of the art algorithms, and they result in enhanced performance on gender and accent recognition. The first part of the thesis focuses on the problem of gender identification, and presents a technique that gives improved performance in situations where training and test conditions are mismatched. The bulk of this thesis is concerned with the application of the i-Vector technique to accent identification, which is the most successful approach to acoustic classification to have emerged in recent years. We show that it is possible to achieve high accuracy accent identification without reliance on transcriptions and without utilising phoneme recognition algorithms. The thesis describes various stages in the development of i-Vector based accent classification that improve the standard approaches usually applied for speaker or language identification, which are insufficient. We demonstrate that very good accent identification performance is possible with acoustic methods by considering different i-Vector projections, frontend parameters, i-Vector configuration parameters, and an optimised fusion of the resulting i-Vector classifiers we can obtain from the same data. We claim to have achieved the best accent identification performance on the test corpus for acoustic methods, with up to 90% identification rate. This performance is even better than previously reported acoustic-phonotactic based systems on the same corpus, and is very close to performance obtained via transcription based accent identification. Finally, we demonstrate that the utilization of our techniques for speech recognition purposes leads to considerably lower word error rates.