Dialect-based speaker classification using speaker-invariant dialect features

In our previous works, a structural pronunciation representation was proposed to extract the linguistic features from dialect pronunciation and classify speakers based on their dialects. In this paper, in order to prove that the structural method can extract the purely speaker-invariant dialectal features, several new experiments are carried out. First, using the data of 19 speakers from different dialect and sub-dialect regions, a dialect-based speaker classification experiment is carried out and satisfactory result is achieved. Then, one Chinese dialectologist transcribes all the data and reads the linguistic content of each original utterance in her voice through looking at the transcript and listening to the original utterance. So a new data set with minimum speaker differences (fixed speaker identity) is created. Using the new data, similar classification experiment is carried out and the result is very similar to the result of last experiment. It means that our method can extract the purely speaker-invariant dialectal features and classify speakers based on their dialects very well. After that, for the original and mimicked data sets, data sets with maximum speaker differences are simulated using high-quality voice morphing techniques. Using the original dialect data and the simulated versions together, classification experiments are carried out based two criteria, spectral comparison and structural comparison. By comparing these results, we can find that unlike the method of spectral comparison, the structural method can purely classify speakers based on their dialects, which shows the proposed dialect structures are speaker-independent and linguistic enough features.