Spectral Subtraction (Vaseghi - Advanced Digital Signal Processing and Noise Reduction), страница 4

However, increasingthe overlap can also increase the correlation of noise frequencies along thetime dimension.351Implementation of Spectral Subtraction% Correct Recognition100with no noise compensationwith spectral subtraction806040200-1001020Signal to Noise Ratio, dBFigure 11.10 The effect of spectral subtraction in improving speech recognition(for a spoken digit data base) in the presence of helicopter noise.11.4.1 Application to Speech Restoration and RecognitionIn speech restoration, the objective is to estimate the instantaneous signalspectrum X(f). The restored magnitude spectrum is combined with the phaseof the noisy signal to form the restored speech signal.

In contrast, speechrecognition systems are more concerned with the restoration of the envelopeof the short-time spectrum than the detailed structure of the spectrum.Averaged values, such as the envelope of a spectrum, can often be estimatedwith more accuracy than the instantaneous values. However, in speechrecognition, as in signal restoration, the processing distortion due to thenegative spectral estimates can cause substantial deterioration inperformance. A careful implementation of spectral subtraction can result ina significant improvement in the recognition performance.Figure 11.9 illustrates the effects of spectral subtraction in restoring asection of a speech signal contaminated with white noise.

Figure 11.10illustrates the improvement that can be obtained from application of spectralsubtraction to recognition of noisy speech contaminated by a helicopternoise. The recognition results were obtained for a hidden Markov modelbased spoken digit recognition.352Spectral Subtraction11.5 SummaryThis chapter began with an introduction to spectral subtraction and itsrelation to Wiener filters.

The main attraction of spectral subtraction is itsrelative simplicity, in that it only requires an estimate of the noise powerspectrum. However, this can also be viewed as a fundamental limitation inthat spectral subtraction does not utilise the statistics and the distributions ofthe signal process. The main problem in spectral subtraction is the presenceof processing distortions caused by the random variations of the noise. Theestimates of the magnitude and power spectral variables, that owing to noisevariations, are negative, have to be mapped into non-negative values.

InSection 11.2, we considered the processing distortions, and illustrated theeffects of rectification of negative estimates on the distribution of the signalspectrum. In Section 11.3, a number of non-linear variants of the spectralsubtraction method were considered. In signal restoration and inapplications of spectral subtraction to speech recognition it is found thatover-subtraction, which is subtracting more than the average noise value,can lead to improved results; if a frequency component is immersed in noisethen over-subtraction can cause further attenuation of the noise.

A formulais proposed in which the over-subtraction factor is made dependent on thenoise variance. As mentioned earlier, the fundamental problem with spectralsubtraction is that it employs relatively too little prior information, and forthis reason it is outperformed by Wiener filters and Bayesian statisticalrestoration methods.BibliographyBOLL S.F (1979) Suppression of Acoustic Noise in Speech Using SpectralSubtraction.

IEEE Tran. on Acoustics, Speech and Signal ProcessingASSP-27, 2, pp. 113–120.BROUTI M., SCHWARTZ R. and MAKHOUL J. (1979) Enhancement of SpeechCorrupted by Acoustic Noise. Proc. IEEE, Int. Conf. on Acoustics,Speech and Signal Processing, ICASSP-79, pp. 208–211.CAPPE O. (1994) Elimination of Musical Noise Phenomenon with theEphraim and Malah Noise Suppressor. IEEE Trans. Speech and AudioProcessing, 2, 2, pp. 345–349.Bibliography353CROZIER P.M. et al (1993) The Use of Linear Prediction and SpectralScaling For Improving Speech Enhancement. EuroSpeech-93, pp. 231234.EPHRAIM Y. (1992) Statistical Model Based Speech Enhancement systems.Proc. IEEE, 80, 10, pp. 1526–1555.EPHRAIM Y. and VAN TREES H.L.

(1993) A Signal Subspace Approach forSpeech Enhancement. Proc. IEEE, Int. Conf. on Acoustics, Speech andSignal Processing, ICASSP-93, pp. 355–58.EPHRAIM Y. and MALAH D. (1984) Speech Enhancement Using a MinimumMean-Square Error Short-Time Amplitude Estimator. IEEE Trans.Acoustics, Speech and Signal Processing.

ASSP-32, 6, pp. 1109–1121.JUANG B.H. and RABINER L.R. (1987) Signal Restoration by SpectralMapping. Proc. IEEE, Int. Conf. on Acoustics. Speech and SignalProcessing, ICASSP-87 Texas.KOBAYASHI T. et al (1993) Speech Recognition Under the Non-StationaryNoise Based on the Noise Hidden Markov Model and SpectralSubtraction. EuroSpeech-93, pp. 833–837.LIM J.S. (1978) Evaluations of Correlation Subtraction Method forEnhancing Speech Degraded by Additive White Noise. IEEE Trans.Acoustics, Speech and Signal Processing, ASSP-26, 5, pp.

471–472.LINHARD K. and KLEMM H. (1997) Noise Reduction with SpectralSubtraction and Median Filtering for Suppression of Musical Tones.Proc. ECSA-NATO Workshop on Robust Speech Recognition, pp.159–162.LOCKWOOD P. and BOUDY J. (1992) Experiments with a Non-linear SpectralSubtractor (NSS) Hidden Markov Models and the Projection, forRobust Speech Recognition in Car, Speech Communications. Elsevier,pp. 215–228.LOCKWOOD P. et al (1992) Non-Linear Spectral Subtraction and HiddenMarkov Models for Robust Speech Recognition in Car NoiseEnvironments. ICASSP-92, pp. 265–268.MILNER B.P. (1995) Speech Recognition in Adverse Environments.

Ph.D.Thesis, University of East Anglia, UK.MCAULAY R.J. and MALPASS M.L. (1980) Speech Enhancement Using ASoft-Decision Noise Suppression Filter. IEEE Trans. ASSP-28, 2, pp.137–145, April.NOLAZCO-FLORES J.A. and YOUNG S.J. (1994) Adapting a HMM-basedRecogniser for Noisy Speech Enhanced by Spectral Subtraction. Proc.IEEE, Int. Conf. on Acoustics, Speech and Signal Processing, ICASSP–94 Adelaide.354Spectral SubtractionPORTER J.E. and BOLL S.F.

(1984) Optimal Estimators for SpectralRestoration of Noisy Speech. Proc. IEEE, Int. Conf. on Acoustics.Speech and Signal Processing, ICASSP-84, pp. 18A.2.1–18A.2.4.O’SHAUGHNESSY D. (1989) Enhancing Speech Degraded by Additive Noiseor Interfering Speakers. IEEE Commun. Mag. pp. 46–52.POLLAK P. et al (1993) Noise Suppression System For A Car.

EuroSpeech93, pp. 1073–1076.SORENSON H.B. (1993) Robust Speaker Independent Speech RecognitionUsing Non-Linear Spectral Subtraction Based IMELDA. EuroSpeech93, pp. 235–238.SONDHI M.M., SCHMIDT C.E. and RABINER R. (1981) Improving the Qualityof a Noisy Speech Signal. Bell Syst. Tech. J., 60, 8, pp. 1847–1859.VAN COMPERNOLLE D.

(1989) Noise Adaptation in a Hidden Markov ModelSpeech Recognition System. Computer Speech and Language, 3, pp.151–167.VASEGHI S.V. and FRAYLING-CORCK R. (1993) Restoration of ArchivedGramophone Records, Journal of Audio Engineering Society.XIE F.(1993) Speech Enhancement by Non-Linear Spectral Estimation aUnifying Approach. EuroSpeech-93, pp. 617–620.ZWICKER E. and FASTEL H. (1999) Psychoacoustics, Facts and Models, 2ndEd. Springer..