Learning vocal mode classifiers from heterogeneous data sources

This paper targets on a generalized vocal mode classifier (speech/singing) that works on audio data from an arbitrary data source. Previous studies on sound classification are commonly based on cross-validation using a single dataset, without considering training-recognition mismatch. In our study, two experimental setups are used: matched training-recognition condition and mismatched training-recognition condition. In the matched condition setup, the classification performance is evaluated using cross-validation on TUT-vocal-2016. In the mismatched setup, the performance is evaluated using seven other datasets for training and TUT-vocal-2016 for testing. The experimental results demonstrate that the classification accuracy is much lower in mismatched condition (69.6%), compared to that in matched condition (95.5%). Various feature normalization methods were tested to improve the performance in the setup of mismatched training-recognition condition. The best performance (96.8%) was obtained using the proposed subdataset-wise normalization.