助听器普通话语音可懂度预测模型

文献信息

中文摘要

目的： 客观语音可懂度预测方法为人工听者测试提供了一种低成本替代方案，可用于研究影响可懂度的多种因素。然而，针对听力受损的普通话听者的语音可懂度客观预测在很大程度上仍不成熟。现有的可懂度指标基于英语语言结构，和/或主要针对正常听力听者设计。本文介绍一种普通话助听器语音感知指数（MCHASPI），这是一种客观指标，用于预测助听器处理后的语音对伴或不伴听力损失的普通话使用者的可懂度。一个关键挑战是高质量数据集的稀缺以及缺乏针对普通话感知特性定制的模型。为解决这一问题，我们首先使用主观听力测试构建了一个普通话可懂度数据集，涵盖六种失真类型的语音。采用听力损失模拟器对参考（未失真）和测试语音信号进行预处理。从这些处理后的信号中提取多种特征并组合用于后续分析。开发了两种预测模型，即MCHASPI-NN和MCHASPI-L，分别利用神经网络和逻辑映射函数从特征值预测可懂度。MCHASPI模型的核心包括一种感知驱动的分段拉伸方案以实现精确的信号对齐，随后进行逐帧阈值检测和多特征融合。对正常听力和听力受损听者的评估表明，在所有测试数据集上，MCHASPI的预测均显著优于现有模型。这验证了MCHASPI在评估不同程度听力损失人群普通话语音可懂度方面的有效性。

英文摘要

Objective speech intelligibility prediction methods provide a low-cost alternative to tests with human listeners, allowing the investigation of many factors affecting intelligibility. However, the objective prediction of speech intelligibility for Mandarin-speaking listeners with hearing impairment is largely underdeveloped. Existing intelligibility metrics are based on English linguistic structures and/or are designed primarily for normal-hearing listeners. This paper introduces a Mandarin Chinese Hearing-Aid Speech Perception Index (MCHASPI), an objective metric for predicting the intelligibility of speech processed by hearing aids for Mandarin Chinese speakers with or without hearing loss. A key challenge was the scarcity of high-quality datasets and the lack of models tailored to the perceptual characteristics of Mandarin Chinese. To address this, we first constructed a Mandarin intelligibility dataset using subjective hearing tests, for speech with six types of distortions. A hearing loss simulator was employed to preprocess the reference (undistorted) and test speech signals. Multiple features were extracted from these processed signals and combined for subsequent analysis. Two prediction models were developed, namely MCHASPI-NN and MCHASPI-L, utilizing a neural network and a logistic mapping function, respectively, to predict intelligibility from feature values. The core of the MCHASPI model involves a perceptually motivated segmental stretching scheme for precise signal alignment, followed by frame-wise threshold detection and multi-feature fusion. Evaluations with normal-hearing and hearing-impaired listeners demonstrated that MCHASPI gave significantly more accurate predictions than existing models across all tested datasets. This validates the effectiveness of MCHASPI for assessing Mandarin speech intelligibility for populations with varying degrees of hearing loss.