Discrete Tokens Exhibit Interlanguage Speech Intelligibility Benefit: an Analytical Study Towards Accent-robust ASR Only with Native Speech Data

TL;DR

This paper investigates accent-robust ASR by testing whether the interlanguage speech intelligibility benefit ($ISIB$) arises in discrete-token based ASR trained solely on native-language data. It treats SSL-derived discrete tokens as proxies for human speech perception and analyzes cross-language tokenization (e.g., English vs. Japanese) using LibriSpeech and accented speech datasets, evaluating performance with metrics $QE$ and $MTER$ alongside standard WER. The results demonstrate ISIB-like effects: native-language tokenization can improve recognition of foreign-accented speech, with stronger gains when the tokenization language matches the speaker's native language, and even showing mismatched ISIB across diverse languages. This suggests that discrete tokens can model perceptual processes and enable accent-robust ASR using only native speech data, broadening applicability to low-resource accents.

Abstract

In this study, we gained insight that contributes to achieving accent-robust ASR using only native speech data. In human perception of non-native speech, the phenomenon known as "interlanguage speech intelligibility benefit" (ISIB) is observed, where non-native listeners who share the native language with the speaker understand the speech better compared even to native listeners. Based on the idea that discrete tokens extracted from self-supervised learning (SSL) models represent the human perception of speech, we conducted an analytical study on the robustness of discrete token-based ASR to non-native speech, varying the language used for training the tokenization, which is viewed as a technical implementation of ISIB. The results showed that ISIB actually occurred in the discrete token-based ASR. Since our approach relies only on native speech data to simulate the behavior of human perception, it is expected to be applicable to a wide range of accents for which speech data is scarce.

Figures (2)

• Figure 1: The reproduction of native speech of language Y by (a) native speakers of language Y, (b) native speakers of language X: the output speech of (b) is expeted to be "X-accented Y".
• Figure 2: The recognition of "X-accented Y" speech by (a) native speakers of language Y, (b) native speakers of language X: (b) is expected to be more accurate through the ISIB effect.