Automatic accent classification of foreign accented Australian English speech

Abstract

English speech based on accent dependent parallel phoneme recognition (PPR) has been developed. The classifier is designed to process continuous speech and to discriminate between native Australian English (AuE) speakers and two migrant speaker groups with foreign accents, whose first languages are Lebanese Arabic (LA) and South Vietnamese (SV). The training of the system can be automated and is novel in that it does not require manually labelled accented data. The test utterances are processed in parallel by three (AuE, SV and LA) accent-specific recognizers incorporating the accent-specific HMMs and phoneme bigram language models to produce accent discrimination likelihood scores. The best average accent classification rates were 85.3% and 76.6% for accent pair and three accent class discrimination tasks, respectively. Analyses of the contributions to accent discrimination by the phoneme level processing, and by the language model, are described.

Automatic classification of speaker characteristics

“…Current research on Australian accent and dialect is focusing on linguistic approach to dialect of phonetic study [13] [14], classification of native and non-native Australian [15], or to improve Australian automatic speech recognition performance [16]. However, there is no research on automatic speaker classification based on the three Australian accents of Broad, General, and Cultivated.…”

On using units trained on foreign data for improved multiple accent speech recognition

“…In many cases, most of the regional variants of a language are handled in a blind way through a global training of the speech recognition system using speech data that cover all regional variants. Accent classification has been studied since many years (Arslan and Hansen, 1996) based either on phone models (Kumpf andKing, 1996, Teixeira et al, 1996) or specific acoustic features (Fung and Liu, 1999). Also, good classification results between regional accents are reported in (Draxler and Burger, 1997) for human listeners on German SpeechDat data, and in (Lin and Simske, 2004) for automatic classification between American and British accents.…”

Emergence of Language-Specific Phoneme Classifiers in Self-Organized Maps

“…They used 9 repetitions of 20 different words to train and test their SOM. Kumpf et al [2] showed that using a Hidden Markov Model (HMM) they were able to classify accents within a group of Australian English speakers with an accuracy of up to 85.3%. Kangas [3] has shown that using a time-dependant representation of Mel-Frequency Cepstral Coefficients (MFCCs) can improve phoneme classification from a 10.4% rate error to a 5.0% rate error.…”