Leveraging Artificial Intelligence and Data Mining to Identify Chinese Dialects: A Comparative Analysis of Tone and Segment Importance

Studies based on big data and deep learning models have been a hot topic in the development of dialect recognition technology. However, the recognition correct rate always fails to reach 100%. To find out the salient features for the recognition of different dialects, a perception experiment was held based on the three dialects of Mandarin Chinese in Liaoning Province. The stimuli consisted of three sets of 30 frequent utterances from Mandarin Speech Perception (MSP) materials. Then, each utterance was manipulated in two ways, i.e., tonally and spectrally. In the tonal dimension the natural fundamental frequency was replaced by a constant f0 of 100 Hz, which removes all melodic information from the utterances. In the second dimension, the original recordings were low-pass filtered using a digital filter with a cut-off frequency of either 1000 Hz or 300 Hz (default smoothing constant of 100 Hz). Next, from each of these dialect areas one male speaker was selected to record the testing materials in their own dialect. And ten eligible subjects were recruited from each dialect region to select the dialects they thought they hear. The results showed that: 1. Segments play a more important role in dialect recognition. 2. The relative contribution of tone varies in different situations. The findings could contribute to further improving the accuracy of dialect recognition technology.