A Semi-Supervised Approach to Automatic Speech Recognition Training For the Icelandic Language

Sammanfattning: Recent advances in deep learning have enabled certain systems to approach or even achieve human parity in certain tasks, including automatic speech recognition. These new state-of-the-art speech recognition models are most often dependent on vast amounts of expensive high-quality labeled speech data for supervised training. In this work, we consider ways of leveraging unlabeled data for unsupervised training to reduce this costly data dependency. Six altered models are compared to a baseline sequence-to-sequence speech recognition model under three different low resource conditions. We show that for all three conditions, a semi-supervised approach surpasses the quality of the baseline.