Implementation of SYMLET Wavelets to Removal of Gaussian Additive Noise from Speech Signal

Transcription

1 Implementation of SYMLET Wavelets to Removal of Gaussian Additive Noise from Speech Signal Abstract: MAHESH S. CHAVAN, * NIKOS MASTORAKIS, MANJUSHA N. CHAVAN, *** M.S. GAIKWAD Department of Electronics Engineering, RIT Sakharale, Dist Sangli (MS), INDIA * Neval Academy of Engineering, BULGERIA **ADCET Ashta, Dist Sangli (MS), INDIA ** Department of Electronics and Telecommunication, SAE Kondwa Pune (MS), INDIA maheshpiyu@gmail.com In hands free speech communication environments situation occurs that speech is superposed by background noise. Over the past few decades there is tremendous increase in the level of ambient environmental noise. This has been due to growth of technology. Noise is added by various factors like noisy engines, heavy machines, pumps, vehicles. Present paper deals with implementation of symlet wavelets for removal of noise in the speech. Details results are depicted in the paper. 1. Introduction There has been a lot of research in speech denoising so far but there always remains room for improvement. Many methods used by researchers so far are either time domain or transform domain. Techniques are based on LPC filtering, Kalman filtering, Hidden markov model, Fourier transform, Discrete cosine transform and simplest one is spectral subtraction. Problem of denoising consist of removing noise from corrupted signal without altering it. The motivation to use wavelet as a possible alternative is to explore new ways to reduce computational complexity and to achieve better noise reduction performance. Wavelet is capable of revealing aspects of data, trends, breakdown points, discontinuities in higher derivatives and self similarity that other techniques miss. Some speech enhancement techniques succeed at improving the quality of the speech without increasing intelligibility. Since speech quality is difficult to quantify, it is not adequate as a measure of success. Objective measures such as signal-to-noise ratios may reflect speech quality, but not intelligibility. There are many more applications for systems that increase intelligibility than for those that only improve quality. Aspects of quality are of course important in reproducing music and song, and, in general, high speech quality is a worthwhile objective. But when speech is subject to noise and other distortions, it is more important to render the speech intelligible than to render it merely more pleasing to the ear. Applications include speech over noisy transmission channels (e.g., in outer space, or over radio waves in bad weather) and speech produced in noisy environments (e.g., in automobiles, aircraft, or outdoor telephone booths)[2-4]. Certain aspects of speech waveforms are more perceptually important than others. The auditory system is more sensitive to the presence of energy than to the absence of it, and tends to ignore many aspects of phase. Thus, speech coding and enhancement algorithms concentrate on accurate preservation of peaks in the speech amplitude spectrum rather than on phase relationships or energy at weaker frequencies. Voiced speech, with its high amplitude and concentration of energy at low frequency, is more perceptually ISBN:

2 important than unvoiced speech for preserving speech quality. Thus, most enhancement algorithms tend to concentrate on improving the periodic portions of speech [1,5-7]. After verifying performance of Haar and Daubechies in this direction now more wavelets such as Symlets and Dmey are implemented and performance is evaluated. 2. Symlet Wavelets: In symn, N is the order. Some authors use 2N instead of N. The symlets are nearly symmetrical, orthogonal and biorthogonal wavelets proposed by Daubechies as modifications to the db family. The properties of the two wavelet families are similar. Here few wavelet are shown in figure 1. number of wavelets as Daubechies. Symlets when applied to signal performs better and SNR of reconstructed or denoised signal is improved. Figure 2 is waveform of when sym6 is applied. From results provided in tables Sym6 is suitable for hard thresholding method. When additive white gaussian noise of 1 db is added to signal it covers whole spectrum of signal so it is difficult to remove. But by observing noisy signal s coefficients and by application of proper threshold noise can be easily removed (a) (b) (c) (d) Figure 2 a) Denoised signal with Sym6 4 spectrogram of denoised signal 35 (e) (f) (g) Figure 1 Symlet wavelet families 3. Implementation of Symlet Wavelet Important wavelet which is found to be efficient in denoising application is Symlet wavelet family. Symlet wavelets used in practice are also selected even b) Spectrogram of denoised signal with Sym6 ISBN:

3 Figure 3 below represents reconstructed signal with level4 decomposition and hard thresholding applied. Wavelet implemented is Sym18. Denoised signal still has some noise. Spectrogram is shown in figure to the right Figure 4 a) Denoised signal with Sym spectrogram Figure 3 a) Denoised signal with Sym spectrogram of denoised signal Figure 4 b) Spectrogram of denoised signal with Sym Figure 3 b): Spectrogram of denoised signal with Sym18 Figure 4 a) indicates denoised signal with Sym18. Signal is denoised using soft thresholding. From spectrogram it is clear that signal has large amount of noise and denoised signal is not similar to original. Results of Symlet Wavelets Results obtained using level4 decomposition for wavelets from Symlet family are discussed below. Results are achieved for different SNR values such as 5, 1, 15 db for additive white gaussian noise and random noise. Hard thresholding results for AWGN shows that Sym1 and Sym18 gives maximum SNR. Sym 1 needs less time as compared to Sym18. Soft thresholding results show that Sym8 have greater SNR than other wavelets. Sym8 is better out of two as it also reconstructs signal within less reconstruction time. Sym8 shows maximum response for random noise hard thresholding also. MSE error for Sym 8 is somewhat higher but can be ignored as it has higher SNR as compared to other having same ISBN:

4 MSE. Sym18 is better in case of soft thresholding of random noise. Reconstruction time is also more for Sym 18. Table 4.1: Results for B1.Wav input file for application of Symlet wavelet Level 4 decomposition with addition of 5dB noise Wavelet Additive White Gaussian noise Random noise Hard thresholding Soft thresholding Hard thresholding Soft thresholding MSE SNR MSE SNR MSE SNR MSE SNR Sym Sym Sym Sym Sym Sym Sym Sym Sym Sym Conclusion: In the paper wavelet based speech denoising algorithm is addressed. Wavelet denoising is a non-parametric estimation method that has been proposed in recent years for speech enhancement applications. In the proposed work all wavelets from Symlet family have higher SNR for level4 than level5 except for Sym6, Sym8, Sym12 and Sym14 in 5 db additive white gaussian noise when hard thresholding is done. For random noise all Symlets have higher SNR values for level4. It is observed that mean square error is approximately same in AWGN for hard and soft thresholding, but somewhat higher in case of random noise for soft thresholding. 5. References 1. Douglas O Shaughnessy, Enhancing Speech Degraded by Additive Noise or Interfering Speakers, February, IEEE Communication magazine. 2. Qin Linmei, Hu Guangrui and Li Chongni, A new Speech Enhancement Method, Proceedings of 21 International Symposium on Intelligent Multimedia, Video and Speech Processing, May 2-4, M. C. E. Rosas-Orea, M. Hernandez- Diaz, V. Alarcon-Aquino, and L. G. ISBN:

5 Guerrero-Ojeda, A Comparative Simulation Study of Wavelet Based Denoising Algorithms, Proceedings of the 15th International Conference on Electronics, Communications and Computers. 4. Yan Long, Lin Gang and Guo Jun, Selection Of The best Wavelet Base For Speech Signal, Proceedings of 24 International Symposium on Intelligent Multimedia, Video and Speech Processing, October 2-22, Saeed Ayat, Mohammad T. Manzuri and Roohollah Dianat, Wavelet based Speech Enhancement Using a new Thresholding Algorithm, Proceedings of 24 International Symposium on Intelligent Multimedia, Video and Speech Processing, October 2-22, Soon Ing Yann Transform based Speech Enhancement Techniques, Soon Ing Yann, PhD Thesis 23, Nanyang Technological University. ISBN: