Speech Quality Perception and its Estimation

语音质量评估是电信、听力学、语音病理学等领域的一个重要课题。例如，语音质量测量用于对电信中的语音编码、降噪和回声抑制算法的性能进行基准测试。同样，语音质量是用户对助听器和辅助听力设备接受和满意的重要因素。同样，语音质量测量在量化语音/声音障碍的严重程度和监测治疗进展方面也很有用。因此，良好的语音质量测量是有效评估通信和听力设备的重要参数，并有助于语音/声音障碍的治疗管理。语音质量可以通过客观或主观手段进行评估。虽然语音质量的主观评价被认为是最可信的标准，但它们也很耗时和昂贵。因此，以合理的精度预测主观语音质量分数的计算方法很有吸引力，也是我们研究工作的重点。在过去的几年里，我们开发了一些新的特征来紧凑地表示语音质量。此外，我们还开发了一个神经模糊特征映射器和一个贝叶斯多元线性样条回归，将这些特征向量映射到语音质量分数中。我们的技术已经成功地应用于电信、听力和语音病理领域的问题。我们计划在未来五年的工作基础上，通过研究更有效的特征提取和特征映射方案，通过纳入听力受损的新模型来预测助听器信号处理算法对感知语音质量的影响，以及通过开发新的工具评估异常语音质量的方法。

Assessment of speech quality is an important topic in a number of fields including telecommunications, Audiology, and Speech Pathology . For example, speech quality measurements are used to benchmark the performance of speech coding, noise reduction and echo suppression algorithms in telecommunications. Similarly, speech quality is an important factor in users' acceptance of and satisfaction with hearing aids and assistive listening devices. Likewise, speech quality measures are useful in quantifying the severity of a speech/voice disorder and in monitoring therapy progress. A good speech quality measure is therefore an important parameter that will allow efficient evaluation of communication and hearing devices, as well as assist in speech/voice disorder therapeutic management.Speech quality can be assessed through objective or subjective means. While subjective opinions ofspeech quality are preferred as the most trustworthy criteria, they are also time-consuming and expensive. Therefore, computational methods that predict subjective speech quality score with reasonable accuracy are attractive and form the focus of our research work.During the past few years, we have developed novel features for compact representation of speech quality. In addition, we have developed a neuro-fuzzy feature mapper and a Bayesian multivariate linear splines regressor to map these feature vectors into speech quality scores. Our techniques have been successfully applied to problems in telecommunications, Audiology, and Speech Pathology fields. We plan to build upon our work during the next five years by investigating more effective feature extraction and feature mapping schemes, byincorporating new models of impaired hearing in predicting the impact of hearing aid signal processing algorithms on perceived speech quality, and by developing new methodologies for instrumental assessment of abnormal voice quality.