An analysis of acoustic features for accented speech classification

Abstract

Spoken language is a topic which lured researchers for a long duration. Due to the variety of different voice-based products, the application of spoken language can be observed in various places. Several home assistant systems have become an integral part of our lives as they make mundane tasks such as setting up reminders and checking emails easy. However, non-native English speakers frequently face problems in using automated assistants because of accented speech. This study presents an analysis of speech accent features for accented speech classification. The aim is to identify which speech features are the most important for accurately classifying accents in spoken language. We collected a dataset of accented speech samples and used various feature extraction techniques to extract relevant features from the speech signal. These features included mel frequency cepstral coefficients, zero-crossing rate, spectral features, chroma features, etc. Machine learning algorithms are used to classify the accents based on the extracted features and achieve an overall accuracy of 86.67%. This research work is prompted by the increasing need to develop robust speech recognition systems that can generalize across regional accents. The performance of standard automatic speech recognition systems drops very often due to accented speech. Several studies tend towards deep learning-based solutions; however, there is a lack of focused analysis of the performance of traditional acoustic features in accent discrimination tasks. This study targets to bridge that gap by performing a comparative study on selected acoustic features. The analysis of speech accent features presented in this study can be useful to develop robust speech accent classification systems for applications such as language learning, speech recognition, and accent identification.