This preview shows pages 1–2. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.View Full Document
Unformatted text preview: MULTI-MICROPHONE NOISE REDUCTION TECHNIQUES FOR HANDS–FREE SPEECH RECOGNITION –A COMPARATIVE STUDY– Joerg Bitzer y , Klaus Uwe Simmer z and Karl-Dirk Kammeyer y y University of Bremen, FB–1, Dept. of Telecommunications P.O. Box 330 440, D–28334 Bremen, Germany, email: email@example.com z Aureca GmbH, Mozartstrasse 26, D–28203 Bremen, Germany, email: firstname.lastname@example.org ABSTRACT In this paper we will describe different multi-microphone noise reduction techniques as for example front-end systems for a command speaker-independent word recognizer in an office environment. Our focus lies on examining the recog- nition rate if the noise source is not gaussian and stationary, but a second speaker in the same room. In this case standard noise reduction techniques like spectral subtraction will fail, and only multi-microphone techniques can raise the recog- nition rate by using the spatial information of the speakers. We will compare the delay-and-sum-beamformer, different superdirective solutions and the post-filter approach. Our results show that these techniques are the right choice for hands–free speech recognition systems. Furthermore, we will give a deeper insight into the superdirective design for microphone arrays. 1. INTRODUCTION More and more users of personal computers work with speech recognition devices to control their systems. The input mi- crophone is almost always headset mounted. This restric- tion is very uncomfortable. To avoid this restriction, hands– free devices are the right choice, but the recognition rate decreases dramatically, as the signal-to-noise ratio (SNR) decreases. Many publications address this problem with the focus on broad-band, slowly varying noise. Single- and multi-microphone approaches are known [1, 2, 3, 4]. This contribution deals with the problem of a second speaker in the same room. Therefore, the interference signal is coloured and non-stationary. For a two-channel system a possible so- lution is published in . This algorithm is a derivation of a two-channel generalized sidelobe canceller . But it can be shown that this kind of algorithms fails in reverberant environments . In this contribution we will focus on non- adaptive solutions. In section two we describe the differ- ent approaches for multi-microphone noise reduction tech- niques. Especially the superdirective design is explained and a design procedure based on the coherence function is given. Section three shows the results of the noise reduction experiments. 2. NOISE REDUCTION ALGORITHMS Noise reduction in a reverberant environment is a difficult problem. We will focus on multi-microphone algorithms which have good capabilities to suppress diffuse noise....
View Full Document
- Spring '10