ISCA Archive Interspeech 2023
ISCA Archive Interspeech 2023

Impact of Residual Noise and Artifacts in Speech Enhancement Errors on Intelligibility of Human and Machine

Shoko Araki, Ayako Yamamoto, Tsubasa Ochiai, Kenichi Arai, Atsunori Ogawa, Tomohiro Nakatani, Toshio Irino

Single-channel (1-ch) speech enhancement (SE) has been widely studied, and high accuracy has been achieved recently. However, enhanced speech still includes some errors that affect human hearing quality and SE applications, e.g., automatic speech recognition (ASR). Previously, [Iwamoto et al., in Interspeech 2022, pp. 5418-5422] decomposed the errors in an enhanced signal into residual noise and artifact components and analyzed their impacts on ASR performance. They showed that the artifacts have a greater impact than the residual noise on ASR. Although the impact on human intelligibility has not been investigated yet, it is essential to get the knowledge to develop SE techniques suitable for both humans and machines. This paper, therefore, investigates the effects of such error factors on human listening. Our subjective test results show that the artifacts have a large impact on human intelligibility, and that residual noise has a lesser impact.