Improving Automatic Speech Recognition for People with Speech Impairment

Abstract - The accuracy of Automatic Speech Recognition (ASR) systems has improved significantly over recent years due to increased computational power of deep learning systems and the availability of large training datasets. Recognition accuracy benchmarks for commercial systems are now as high as 95% for many (mostly typical) speakers and some applications. Despite these improvements, however, recognition accuracy of non-typical and especially disordered speech is still unacceptably low, rendering the technology unusable for the many speakers who could benefit the most from this technology.

Google’s Project Euphonia aims at helping people with atypical speech be better understood. I will give an overview to our large-scale data collection initiative, and present our research on both effective and efficient adaptation of standard-speech ASR models to work well for a large variety and severity of speech impairments.

Short bio - Katrin earned her Ph.D. from University of Dortmund, supervised by Prof Katharina Morik and Prof Udo Hahn (FSU Jena), in 2010. She has since worked on a variety of NLP, Text Mining, and Speech Processing projects, including eg Automated Publication Classification and Keywording for the German National Library, Large-Scale Patent Classification for the European Patent Office, Sentiment Analysis and Recommender Systems at OpenTable, Neural Machine Translation at Google Translate. Since 2019, Katrin leads the research efforts on Automated Speech Recognition for impaired speech within Project Euphonia, an AI4SG initiative within Google Research.

In her free time, Katrin can be found exploring the beautiful outdoors of the Bay Area by bike or kayak.