Petr Pollak: List of on-line avialable publications:
The Noise Suppression System for a Car.
Eurospeech'93, Berlin.
Cepstral Speech/Pause Detectors.
IEEE NSIP'95, Chalkidiki.
The Study of Speech/Pause Detectors for Speech Enhancement Methods.
Eurospeech'95, Madrid.
Extended Spectral Subtraction.
Eusipco'96, Trieste.
Study of Speech Recognition in Noisy Environment.
ECSAP'97, Prague.
BACK TO:
Publication List (Petr Pollak)
|
Home page (Petr Pollak)
|
Speech Processing Group (K331)
|
Speech Enhancement Group (K331)
Problems or comments - mail to: pollak@feld.cvut.cz.
I will appreciate any information about Your experiences which You have made with presented algorithms or problem solutions.
Last change 3 Oct 1998
THE NOISE SUPPRESSION SYSTEM FOR A CAR.
Petr Pollak
& Pavel Sovka
& Jan Uhlir.
In proc. of the 3rd European Conference on Speech Communication and Technology
- EUROSPEECH'93, pp.1073-1076, Berlin, Germany, Sep 1993.
ABSTRACT
The whole system for noise suppression in speech recorded in
a running car was designed. One channel spectral subtraction
method with full-wave rectification was chosen because of its
robustness, simplicity, and non-musical tone output. The
improvement of noise suppression was gained by the repetition
of this method. Directional microphones for the signal picking up
were chosen to improve the input signal-to-noise ratio (SNR)
of corrupted speech signal.
Segment speech/pause detector based on energy tracking was used
with some prefiltration of corrupted speech to improve detector
function.
Scanned version of proceeding paper in PDF format
CEPSTRAL SPEECH/PAUSE DETECTORS.
Petr Pollak
& Pavel Sovka
In proc. of 1995 IEEE Workshop on Nonlinear Signal and Image Processing,
pp.388-391, Neos Marmaras, Halkidiki, Greece, June 20-22, 1995.
ABSTRACT
Many systems for noisy speech processing usually require reliable
speech/pause detector.
This paper describes two algorithms for speech/pause
cepstral detectors. Integral cepstral algorithm and
differential algorithm based on differenced cepstrum.
Both algorithms use smoothing procedure based on median
filtering. New criteria have been used for detectors
comparisons. Many experiments confirmed detectors
reliability and their ability to detect speech in
real car noise with high probability.
The computational cost of presented algorithms is slow so they are
suitable for real-time implementation.
Scanned version of the paper in the PDF format
THE STUDY OF SPEECH/PAUSE DETECTORS FOR SPEECH ENHANCEMENT
METHODS.
Sovka,P.
& Pollak,P.
Proceedings of the 4th European
Conference on Speech Communication and Technology,
pp.1575-1578, Madrid, Spain, September 1995.
ABSTRACT
Speech/pause detectors are the limiting parts of systems for the
suppression of additive noises in speech, because the quality of the
detector determines the performance of the whole noise suppression
system. If the speech/pause decision is not correct then speech echoes
and residual noises are present in enhanced speech. Information
about speech activity is need not only for an estimation of
background noise characteristics but also for time delay
compensation of signals picked up by microphone array.
Basic principles of various adaptive algorithms for speech detection
in a noise and their behaviour under real car noise
conditions are described.
Energy, spectral, cepstral, and coherence detectors are compared.
All these algorithms are suitable for real time
implementation with one or two microphones. High probability
of correct speech/pause detection can be obtained even if signal
to noise ratio is low and noises are highly nonstationary.
Scanned paper in the PDF format
EXTENDED SPECTRAL SUBTRACTION
Sovka,P.
& Pollak,P.
& Kybic,J.
In proc. of European Conference on Signal processing and Communication,
Trieste, September, 1996.
ABSTRACT
The spectral subtraction offers the simple and computationally
efficient tool for the suppression of an additive noise in a speech
signal. This method has been extensively studied for almost
twenty years. The research has been focused on higher degree
of noise suppression, lower speech distortion, and less
audible musical noise. The last requirement is
important especially in the hand-free telephony application. But
the main shortcoming of this method has not been overcome for a
long time. It is the updating of the background noise
characteristics estimation, especially during speech sequences.
The extended spectral subtraction overcomes the typical disadvantage
of the standard spectral subtraction technique - the impossibility of noise estimation
during speech sequence. Our method is the combination
of Wiener filtering and spectral subtraction. The noise can be succesfully
updated even during the speech sequences and that is why there is no need of
voice activity detector.
Postscript version of the paper
Paper in PDF-format
STUDY OF SPEECH RECOGNITION IN NOISY ENVIRONMENT
Kreisinger,T.
&
Pollak,P.
&
Sovka,P.
&
Uhlir,J.
In proc. of 1-st European Conference on Signal prediction and Analysis,
Prague, June, 1997.
ABSTRACT
This paper addresses effects of mismatched conditions and
their minimization with respect to the performance
of speaker-independent isolated
word recognition in the car-noise environment
without consideration of Lombard effect.
This study is primarily intended to study the dependence of the
recognition rate on the SNR of an input signal without and with
noise enhancement preprocessing, especially to find conditions
under that the
modified spectral subtraction can be effectively used for the
speech recognition in a real non-stationary car-noise environment.
If as the worst recognition rate is admitted e.g. 80\%, then the
use of spectral subtraction methods enables to use wider interval
of input SNRs:
for the trainig made on a clean speech this interval is (40,6) dB;
for the training made on a noisy speech this interval is (40,-2) dB;
for the training performed on an enhanced speech this interval
is (40, -8) dB. The third case gives the widest interval of SNRs in which a
recogniser (with the final recognition rate in the interval
of (100,80)\%) can be used.
Postscript version of the paper - It will be here soon !