Subjective intelligibility estimation of noisy speech using objective measures

2015 Fiscal Year Final Research Report

• Project/Area Number: 25330182
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Perceptual information processing
• Research Institution: Yamagata University
• Principal Investigator: Kondo Kazuhiro 山形大学, 理工学研究科, 教授 (10312753)
• Research Collaborator: KOBAYASHI Yosuke 室蘭工業大学, 大学院工学研究科, 助教 (10735103)
• Project Period (FY): 2013-04-01 – 2016-03-31
• Keywords: 音声了解度 / 客観推定 / 機械学習 / クラスタリング

Outline of Final Research Achievements

We investigated a method to estimate speech intelligibility of noisy speech using objective measures without human listeners.
We investigated clustering noise according to its effect on intelligibility, and using separate mapping function from objective measure to subjective intelligibility by cluster. We found that three clusters for this purpose is appropriate. By training a relatively simple mapping function per cluster, the estimation accuracy improved significantly compared to using a single function for all noise. The accuracy improved further by using support vector regression, a well known method in machine learning, as the mapping function. We also investigated on estimation without the use of clean speech samples, and found that accuracy high enough for practical use is possible.

Free Research Field

情報学