Pre-trained Speaker recognition embedding models for disorder speech classifications

This project aims to investigate the binary classification of voice disorders using speaker verification embedding models. The embeddings of speech samples from Hungarian and Dutch samples were extracted using two state-of-the-art speaker verification models, particularly in cross-lingual dysphonic speech detection. The severity level of dysphonic speech was then estimated. To test the algorithms' performance in cross-lingual scenarios, speech samples from one language will be used for training and another for testing. Generating new samples or resamples from the existing samples using the bootstrapping method to evaluate the precision of a sample statistic. Boosting is an ensemble method that involves successively training many homogeneous algorithms. A final model with the greatest outcomes is produced by these distinct algorithms. The method generates new hypothetical samples that aid in testing an estimated value using the replacement technique.