Excitation Source Feature Based Dialect Identification in Ao — A Low Resource Language

Ao is an under-resourced Tibeto-Burman tonal language spoken in Nagaland, India. There are three distinct dialects of the language, namely, Chungli, Mongsen and Changki. The objective of dialect identification is to identify one dialect from the other within the same language family. The goal of this study is to ascertain the potential of excitation source features for automatic dialect identification in Ao. In this direction, Integrated Linear Prediction Residual (ILPR), an approximate representation of source signal, is explored. The log Mel spectrogram of ILPR (SExt) signal is used to exploit the time-frequency characteristics of the excitation source. This work proposes attention based CNN-BiGRU architecture for automatic dialect identification tasks. Additionally, log Mel spectrogram (SVT), extracted from the pre-emphasized speech signal, is used as a baseline method. The (SVT) contains the vocal-tract characteristics of the speech signal. A significant performance improvement of (nearly) 6% accuracy is observed when the excitation source feature (SExt) is combined with the vocal tract representation (SVT). To analyse the effect of segment duration, dialect identification performance is reported for three different durations, viz., 1 sec, 3 sec and 6 sec. The effect of gender in dialect identification task for Ao is also studied in this work.