Low-Resource Dialect Identification in Ao Using Noise Robust Mean Hilbert Envelope Coefficients

Abstract

This paper presents an automatic dialect identification system in Ao using a deep Convolutional Neural Network with residual connections. Ao is an under-resourced language belonging to the Tibeto-Burman family in the North-East of India. The three distinct dialects of the language are Chungli, Mongsen and Changki. Ao is a tone language and consists of three tones, viz., high, mid, and low. The recognition of tones is said to be influenced by the production process as well as human perception. In this work, the Mean Hilbert Envelope Coefficients (MHEC) feature is explored to identify the three dialects of Ao as this feature is reported to have information of human auditory nerve responses. Mel Frequency Cepstral Coefficients (MFCC) feature is used as the baseline. In addition, the effect of noise in the dialect identification task at various signal-to-noise ratio scenarios is studied. The experiments show that the MHEC feature provides an improvement of almost 10% average F1-score at high noise cases.