Key Laboratory of Noise and Vibration Research, Institute of Acoustics, Chinese Academy of Sciences, 100190, Beijing, China.
Cambridge Hearing Group, Department of Psychology, University of Cambridge, Downing Street, Cambridge CB2 3EB, United Kingdom.
J Acoust Soc Am. 2024 Nov 1;156(5):3088-3101. doi: 10.1121/10.0034233.
Objective indices for predicting speech intelligibility offer a quick and convenient alternative to behavioral measures of speech intelligibility. However, most such indices are designed for a specific language, such as English, and they do not take adequate account of tonal information in speech when applied to languages like Mandarin Chinese (hereafter called Mandarin) for which the patterns of fundamental frequency (F0) variation play an important role in distinguishing speech sounds with similar phonetic content. To address this, two experiments with normal-hearing listeners were conducted examining: (1) The impact of manipulations of tonal information on the intelligibility of Mandarin sentences presented in speech-shaped noise (SSN) at several signal-to-noise ratios (SNRs); (2) The intelligibility of Mandarin sentences with intact tonal information presented in SSN, pink noise, and babble at several SNRs. The outcomes were not correctly predicted by the Hearing Aid Speech Perception Index (HASPI-V1). A new intelligibility metric was developed that used one acoustic feature from HASPI-V1 plus Hilbert time envelope and temporal fine structure information from multiple frequency bands. For the new metric, the Pearson correlation between obtained and predicted intelligibility was 0.923 and the root mean square error was 0.119. The new metric provides a potential tool for evaluating Mandarin intelligibility.
客观的语音可懂度预测指标为语音可懂度的行为测量提供了一种快速而方便的替代方法。然而，大多数这样的指标都是为特定的语言设计的，例如英语，并且当应用于普通话等语言时，它们没有充分考虑语音中的声调信息，而在普通话中，基频（F0）变化的模式在区分具有相似语音内容的语音方面起着重要作用。为了解决这个问题，进行了两项正常听力受试者的实验，考察了：（1）声调信息的操纵对在几种信噪比（SNR）下的语音噪声（SSN）中呈现的普通话句子的可懂度的影响；（2）在 SSN、粉红噪声和噪声中呈现具有完整声调信息的普通话句子的可懂度在几种 SNR 下的情况。这些结果不能被听力助听感知指数（HASPI-V1）正确预测。开发了一种新的可懂度度量标准，该标准使用了 HASPI-V1 的一个声学特征，加上来自多个频带的希尔伯特时间包络和时间精细结构信息。对于新的度量标准，获得的和预测的可懂度之间的皮尔逊相关系数为 0.923，均方根误差为 0.119。新的度量标准为评估普通话可懂度提供了一种潜在的工具。
