NOAA-LSTM: A New Method of Dialect Identification

Abstract

Dialect Identification (DID) is a particular case of general Language Identification (LID). Due to the high similarity of dialects and similar phonetic features in adjacent areas, DID is a more challenging problem. Long Short-Term Memory (LSTM) networks tend to be used and do well in LID tasks in recent years, but do not have a good performance on DID tasks. In this paper, NOAA (New One-against-all) binary classifier based on OAA (One-against-all) binary classifier obtained proposed, and a new dialect recognition method was combining NOAA with LSTM networks was offered under the guidance of Chinese humanities. The new approach achieves better performance on a DID task than a single LSTM network. The experiment was conducted based on six major dialects in China, and trained under the acoustic characteristics of log Mel-scale filter banks energies (FBANK). Experimental results on six dialects recognition tasks indicate that the accuracy of the new method is higher than that of a single LSTM network.