My main linguistic interests are scattered wherever I can find relations to speech. I am especially interested in how and why computer-workable representations of natural language data are structured the way they are.
Uses torch to probe transformer-based audio encoder embeddings, over the TIMIT dataset. The hope is to develop our human intuition for how audio encoder models represent human understanding of acoustic features in speech data.
A deep learning-based disfluency detector for a seminar at SJSU. We extend work by Liu et al.(2023) and Mohapatra et al.(2022) for experimenting with multi-encoder disfluency detection using wav2vec, voc2vec, and Whisper. Trained on SEP-28k.
A tool using Python and librosa to generate spectrograms for recordings of Silbo Gomero, a whistled language from the island of La Gomera.