Accented Speech Recognition: A Survey

Abstract

Automatic Speech Recognition (ASR) systems generalize poorly on accented speech. The phonetic and linguistic variability of accents present hard challenges for ASR systems today in both data collection and modeling strategies. The resulting bias in ASR performance across accents comes at a cost to both users and providers of ASR.We present a survey of current promising approaches to accented speech recognition and highlight the key challenges in the space. Approaches mostly focus on single model generalization and accent feature engineering. Among the challenges, lack of a standard benchmark makes research and comparison especially difficult.

Accented Speech Recognition: Benchmarking, Pre-training, and Diverse Data

“…Modeling approaches include accent conversion [14], where a transformation is applied to a non-native utterance to make it sound as if the speaker had a native accent. For a thorough overview of research on improving speech recognition for accents, see [15].…”

Bridging the Gap: Using Deep Acoustic Representations to Learn Grounded Language from Percepts and Raw Speech

“…These large datasets are usually derived from fairly heterogeneous groups (Koenecke et al 2020). Given this, we see gender (Alsharhan and Ramsay 2020;Tatman 2017), race (Blodgett and O'Connor 2017), disability status (Fok et al 2018), and native language/dialect (Hinsvark et al 2021) disparities in successful ASR, reducing technology accessibility for underrepresented groups. Contemporaneous work (Liu et al 2021) introduced a dataset to measure ASR performance across age, gender and skin type.…”

“…In this analysis, we split based on perceived gender, accent, muffled-ness, background noise, and volume. The accessibility hurdles faced by members of minority populations in learning systems are well documented (Hinsvark et al 2021;Tatman 2017;Koenecke et al 2020;Alsharhan and Ramsay 2020). These hurdles can be attributed in part to lack of representation of minority groups in large datasets, but other factors also come into play, especially with smaller datasets and feature-engineered methods.…”

“…These systems are rarely developed in tandem with the robotics community, and so do not take the unique challenges of robotic sensing into account (Marge et al 2022). In addition, current ASR systems work inconsistently across demographics (Tatman 2017;Hinsvark et al 2021), which represents a problem in inclusive design. They are also "black box" systems that cannot improve their speech recognition from their environment or other perceptual clues.…”

The pitfalls of negative data bias for the T-cell epitope specificity challenge

“…Similar issues exist within many other fields. For example for speech recognition, a big challenge is the low variability of dialects and accents in the available data 14 . An example with a lack of negative data is anomaly detection, where rare events by definition only occur infrequently 15 .…”