Speech Science Forum - Odette Scharenborg (TU Delft)

Abstract:

Automatic speech recognition (ASR) is increasingly used, e.g. in emergency response centres, domestic voice assistants (e.g., Google Home), and search engines. They work (really) well for standard speakers of languages for which there is enough training data to train the systems. However, practice and recent evidence suggests that state-of-the-art ASRs struggle with the large variation in speech due to e.g., gender, age, speech impairment, race, and accents, i.e., ASR systems tend to not work well for "non-standard" speakers of the language.​

In this talk, I will present our recent research on inclusive automatic speech recognition, i.e., ASR systems that work for everyone, irrespective of how someone speaks or the language that person speaks. The overarching goal in our new research line on inclusive automatic speech recognition is to uncover bias in ASR systems in order to work towards proactive bias reduction in ASR. In this talk, I will provide an overview of possible factors that can cause this bias; I will present systematic experiments aimed at quantifying and locating the bias of state-of-the-art ASRs on speech from different groups of speakers; and I will present recent research on reducing the bias against non-native accented Dutch. In our work, we focus on bias against gender, age, regional accents and non-native accents.​