Site menu:

Experimental results of a multi-channel speech dereverberation algorithm based on a statistical model of late reverberation

E.A.P. Habets

Abstract

In general, acoustic signals radiated within a room are linearly distorted by reflections from walls and other objects. These distortions degrade the fidelity and intelligibility of speech and the recognition performance of automatic speech recognition systems. We have investigated the application of signal processing techniques to the improvement of the quality of speech distorted in an acoustic environment.

One important effect of reverberation on speech is overlap–masking, i.e. previous phonemes overlap following phonemes. In [1] a multi–channel speech dereverberation method based on Spectral Subtraction was introduced to reduce this effect. The described method estimates the instantaneous power spectrum of the reverberation based on Polack’s statistical model of late reverberation.

In this paper we present experimental results obtained using signals measured in a real acoustic environment. Additionally we proposed a modified gain function to reduce noise.

References

  1. E.A.P. Habets [details + audio examples]
    Multi-channel speech dereverberation based on a statistical model of late reverberation, Proc. of the 30th IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2005), Philadelphia, USA, Mar. 18-23, 2005, pp. 173-176.

Results

Fig 1. Microphone Array

We have used an uniform linear array of 5 omni-directional microphones (see Figure 1), the inner-microphone spacing was 10 cm. The source-receiver distance was 3 m. All signals were measured in our audio lab. The reverberation time was 296 ms. The source signal consists of a clean female speech fragment of 8.5 seconds, sampled at 8 kHz. In this experiment all reflections arriving later than T ms (40 ms or 60 ms) after the arrival time of the direct wave are assumed to be late reverberant and will be reduced by the proposed algorithm.

Clean
Reverberant
Delay & Sum Beamformer
Proposed T=40 ms
Proposed T=60 ms

Note: It is strongly recommended to use headphones for listening tests.

Status

Published in the Proc. of the first annual IEEE BENELUX/DSP Valley Signal Processing Symposium (SPS-DARTS 2005), Antwerp, Belgium, Apr., 2005, pp. 11-14.

BibTex Entry

@INPROCEEDINGS{Habets2005,
  author = {E. A. P. Habets},
  title = {Experimental Results of a Multi-Channel Speech Dereverberation Algorithm
	based on a Statistical Model of Late Reverberation},
  booktitle = {Proc. first annual {IEEE} {BENELUX}/{DSP} Valley Signal Processing
	Symposium ({SPS}-{DARTS})},
  year = {2005},
  pages = {11--14},
  address = {Antwerpen, Belgium},
  month = apr
}

[Back]