Semi-Supervised Learning for Acoustic and Prosodic Modeling in Speech Recognition

Опубликовано 27 июля 2016, 23:25

Semi-supervised learning is a class of machine learning techniques that aims to use unlabeled data to improve the performance of models trained by labeled data only. It is especially useful in scenarios where enormous amounts of unlabeled data are available with little cost. In this work, I investigate semi-supervised learning algorithms for prosodic and acoustic modeling. In particular, I propose an integrated training framework, which utilizes untranscribed speech to discover prosodic boundaries, or to improve phonetic classification and recognition accuracy. We show promising results for Mandarin prosodic boundary detection, and significant improvement for phonetic classification.