权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Study on Adaptive Comb filter for Auditory Scene Analysis

听觉场景分析的自适应梳状滤波器研究

基本信息

批准号：
12650420
负责人：
NISHI Kazuki
金额：
$ 1.66万
依托单位：
The University of Electro-Communications
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (C)
财政年份：
2000
资助国家：
日本
起止时间：
2000 至 2001
项目状态：
已结题

来源：
https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-12650420/
关键词：
Quasi-periodic Signal comb filter Auditory Scene Analysis Wiener filter kalman filter Stochastic differential equation constant-Q Critical Bands 櫛形フィルタウィーナーフィルタ振幅変動ピッチ変動

项目摘要

In this study, we have developed optimal filters for quasi-periodic signals involving amplitude and pitch fluctuations (e.g., a real speech or instrumental sound) as a fundamental technology for realization of the auditory scene analysis.First, we derived linear optimal filters based on the quasi-periodic signal model. The Wiener filter solution under the steady-state condition leads to the constant-BW (-bandwidth)/constant-Q mixed comb structure in which each pass-band located at harmonic frequencies has a constant bandwidth in lower frequency, constant-Q characteristic in higher frequency, and those mixture appears in the intermediate frequency. It is shown that such filter banks closely imitates that of the human auditory system. Next, we derived a Kalman filter based on the Ito stochastic differential equation involving unknown random fluctuations as well as the given slowly varying fluctuations. It is verified that the steady-state solution coincides with the Wiener filter one.In addition, we have designed digital filters, which have not only a comb structure for extracting each harmonic component but also a notch structure for suppressing interference between harmonics components, based on digitization of the continuous Kalman filter equation. Finally, we have established a filter system for extracting a specific sound from among the mixed sound, and confirmed the validity of utilizing as the front-end processing for speech recognition.

在这项研究中，我们针对包含幅度和音调波动的准周期信号(例如，真实语音或乐器声音)开发了最优滤波器，作为实现听觉场景分析的基础技术。稳态条件下的维纳滤波解导致了恒定带宽/恒定Q混合梳状结构，其中每个通带位于谐波频率处在低频具有恒定带宽，在高频具有恒定Q特性，而这些混合出现在中频。结果表明，这种滤波器组与人类听觉系统的滤波器组非常接近。接下来，我们基于包含未知随机涨落和给定的缓变涨落的Ito随机微分方程，推导出卡尔曼滤波。此外，在连续卡尔曼滤波方程数字化的基础上，设计了既能提取各次谐波分量的梳状结构又能抑制各次谐波分量间干扰的陷波结构的数字滤波器。最后，我们建立了一个从混合声音中提取特定声音的过滤系统，并验证了将其作为语音识别的前端处理的有效性。