Realization and Performance Evaluation of New Hybrid Speech Compression Technique

Transcription

1 Realization and Performance Evaluation of New Hybrid Speech Compression Technique Javaid A. Sheikh Post Graduate Department of Electronics & IT University of Kashmir Srinagar, India Sakeena Akhtar Post Graduate Department of Electronics & IT University of Kashmir Srinagar, India Abstract: A new technique for speech compression using combination of Linear Predictive Coding (LPC) with transform coding has been proposed and evaluated in this paper. The conventional compression technique based on LPC has been replaced with two hybrid models; LPC and Discrete Cosine Transform (DCT), LPC and Discrete Wavelet Transform (DWT). Three different speech signals are compressed by linear prediction, DCT and DWT. We compare the performance of different speech compression techniques on test recorded speech signals. It has been found that hybrid technique of Linear Prediction and Discrete Wavelet Transform outperforms other techniques used in the paper. The proposed scheme has lot of significance in wireless communication in general and mobile communications in particular. Keywords: Linear Predictive Coding, Discrete Cosine Transform, Discrete Wavelet Transform, Peak Signal to Noise Ration, Compression Ratio I. INTRODUCTION Speech is a naturally occurring signal and hence random in nature and is the most common way of communication for humans. Speech compression is to remove redundancy from the speech signal to achieve compression and to reduce transmission (e.g., for voice transmission over mobile channels with limited capacity) or to reduce the storage costs (e.g., for speech recording). Speech compression changes input speech data stream into a reduced size data stream, by removing inherent redundancy associated with speech signals. The compression of speech has found a great importance in communication systems. As in wireless communication, bandwidth is a major resource that is taken into consideration and the service providers are often met with the challenge of accommodating more number of users within a limited bandwidth. Compression of speech can overcome this challenge by enhancing the bit rate and providing a good speech quality. By using Compression techniques, the program execution time and storage of processor are reduced to a greater extent. Compression reduces the data transfer rate and bandwidth requirement with security of data [1]. Compression techniques are classified into two broad categories: lossless and lossy. In Lossless compression, the actual signal is obtained from the compressed file. In case of Lossy compression, the actual file cannot be completely obtained from the compressed file. Speech compression is a type of lossy compression that means the output speech signal is not exactly same as that of the input speech signal [2]. II. RELATED WORK M. V. Patil et-al. (2013) presented a paper on various transform coding techniques. A simple Discrete Wavelet Transform and Discrete Cosine Transform based audio compression scheme is being presented and it is observed that with DWT technique, the performance parameters in terms of Peak Signal to Noise Ratio and Compression Ratio of a speech signal show improved results and it is also being observed that specific wavelets have varying effects on a speech signal being represented [3]. Siva Nagu. T et-al. (2012) proposed speech compression based on Wavelet Coding and Adaptive Kalman Filtering. A speech signal is compressed using Wavelet Coding and then with Adaptive Kalman Filtering and it is observed that the signal being compressed with Wavelet Coding and Kalman Filtering shows better performance [4]. Harmanpreet Kaur et-al. (2012) highlighted the key features of various wavelet filters used for speech processing. It is found that a signal compressed with wavelet filters show improved fidelity and efficient compression in terms of Mean Square Error and Peak Signal to Noise Ratio [5]. Satish Kumar et-al. (2012) presented a paper on Wavelet Compression Techniques. It is being observed that with Wavelet Compression Technique the ratio of compression and signal to noise of a speech signal can be well-adjusted. Haar and db6 wavelet are used to decompose a signal at level 6 [6]. All of the above mentioned techniques [3, 4, 5 and 6] have either used DCT or DWT compression techniques, none of them have discussed about combining these two techniques or combining them with any other compression technique. As LPC is the prominent speech coding technique, thus in our proposed work an attempt has been made to combine the DWT or DCT features with the predictive analysis of speech so that improved performance of the compressed speech signal is obtained. The proposed techniques combined the features of both transform coding and predictive coding. III. TYPES OF SPEECH CODING TECHNIQUES Three basic speech compression techniques are commonly used in practice namely: Waveform, Predictive and Transform coding. Waveform based speech compression is mainly used to eliminate redundancy in the speech waveform and to reconstruct the speech waveform at the decoder side as closely as possible to the original speech waveform. Waveform-based RES Publication 2012 Page 16

2 speech compression techniques are simple and less complex but low compression. The standard bit rate for most of the waveform speech compression is 64kbps to 16kbps. The bit rate lower than the 16kbps yields higher quantisation error and this results in lower speech quality. Typical waveform-based speech compression techniques are PCM (µ-law and A-law) and ADPCM (Adaptive Differential PCM). Parametric-based speech compression technique is based on the features that speech signal is static or the shape of the vocal tract is unchanging in short period of time (e.g., 10 ms). During this period of time, a speech segment can be categorized as either a voiced or unvoiced speech segment. For each speech segment, the vocal tract filter constraints, voiced/unvoiced decision, pitch period and gain (signal energy) parameters are obtained via speech analysis at the encoder. These constraints are then coded into binary bit stream and sent to transmission channel. The decoder at the receiver side will reconstruct the speech (carry out speech synthesis) based on the received parameters. Compared to waveform-based codecs, parametricbased codecs are highly difficult to implement but has better compression ratio. The quality of parametric based speech codecs is low, with mechanic sound, but with reasonable intelligibility. A typical parametric coding technique is Linear Predictive Coding (LPC) which has a bit rate from 1.2 to 4.8 kb/s and is normally used in secure wireless communications systems when transmission bandwidth sufficiently low. As parametric-based codecs cannot achieve high speech quality because of the use of simple classification of speech segments into either voiced or unvoiced speech and simple representation of voiced speech with impulse period train, Transform coding techniques were proposed to combine the features of both waveform-based and parametric-based coding. It keeps the nature of parametric coding which includes vocal tract filter, pitch period analysis and voiced/unvoiced decision. Instead of using an impulse period train to represent the excitation signal for voiced speech segment, it uses waveformlike excitation signal for voiced, unvoiced or transition (containing both voiced or unvoiced) speech segments. Transform coding technique includes FFT, DCT, CWT and DWT [7]. In this paper only predictive coding and transform coding have been used. A. Linear Predictive Coding Linear predictive coding is a parametric based speech compression in which spectral envelope of a speech signal is represented in compressed form using the information of linear predictive model. The basic idea is to predict current sample from the past samples. How many past samples we need depends upon the pole model. (If we use 12 pole models, then 12 past samples are used). We can estimate vocal tract model to predict the filter coefficients. The transform function of vocal tract model is given as: where s(z) is the output speech u(z) is the excitation G is a by-product ak is the filter coefficients (every 10ms 12 coefficients will come out) Taking Inverse Fourier Transform, we get or since Gu(n)=0; between pitch pulses excitation is zero therefore, where s(n) is the current sample s(n-k) is the past sample ai is the predictor coefficients. Gu(n) is the excitation LPC is often referred to as inverse filtering as it determines the all zero filter which is inverse of the vocal tract model (IIR Filter). The practical values of P = 9 to 15. For P = 9: 4 formant resonance frequencies are required, for P = 15: 7 formant resonance frequencies are required and every formant resonance is actually modelled by cascading two digital formant filter (both are digital filters with one pole). The fundamental idea is that for a speech signals having a bit rate of 2.48kbps, the speech signal is assumed to be sampled at 8000 samples per second. It is broken in segments of size 180 samples each, and is first passed through a low-pass filter using of bandwidth 1 khz. The energy at the output of the filter is used to take the voiced/unvoiced decision for the segment. The estimate of energy in the background noise is the energy of the unvoiced segment. The decision is further refined by calculating the zero crossing rate. For pitch period measurement, the LPC-10 standard uses average magnitude difference function (ADMF). LPC-10 uses the co-variance method for the computation of reflection coefficients. Thus the parameters transmitted to the receiver include voicing decision, pitch period, and vocal tract parameters namely, reflection coefficients. One bit is used for the voicing decision. The pitch is quantized to 1 of 60 different values using a log companded quantizer. The algorithm uses a 10th order filter for the voiced segment and a 4th order filter for the unvoiced segment [8]. B. Transform Coding In transform coding technique, the signal is transformed (1) (2) (3) (4) RES Publication 2012 Page 17

3 into frequency domain. In this paper Discrete Cosine Transform technique and Discrete Wavelet Transform technique are being used. The energy compaction property of DCT is very important for speech coding as a few transform coefficients (called as DCT coefficients) are used to represent the majority of energy in sequence. This attribute of DCT reduces the data. DCT can be used for compression of speech signals due to high correlation between the adjacent coefficients. a sequence can be reconstructed very accurately using a very few DCT coefficient. It is because of this property of DCT, effective data compression can be achieved. DCT of 1D sequence x(n) of length N is given by: where j=0,1,2,3..n-1 Inverse DCT of the sequence is given as: where n=0,1,2,3.n-1 for equations (5) and (6) n putting j=0 in equation (5), we get First transformed coefficient is the average value of the sample sequence and is referred as the DC confidents. The other transform coefficients are known as AC coefficients. Thus finite sequence of data points are thus expressed in terms of sum of cosine function oscillating at different frequencies and are very common coding technique for speech signals. Wavelet is a new technique for examining and comparing a speech signal, it is more advantageous technique because it holds both time and frequency aspect of a signal. Wavelet breaks speech signal into different coefficients. Some of the coefficients having small value are treated as insignificant during data compression and are hence discarded [9]. Wavelets are obtained by a single Mother Wavelet by delay and shifting. (9) Where a and b are called as scaling and shifting parameters. DWT due to its orthogonal properties do not produce any redundant information. The scaling and shifting parameters are thus responsible for Multiresolution Analysis Algorithm, which decomposes a signal into scales having different time and frequency resolutions. Various Mother wavelet (e.g. Haar, Daubechies, Coiflets, Symlet, Biorthogonal and etc.) functions are differentiated upon these scaling functions, thus the choice of wavelet decides the final (5) (6) (7) (8) waveform shape. In our proposed work Daubechies wavelet db4 has been implemented. IV. SIGNAL DECOMPOSITION In DWT, a signal (S(t)) to be examined is passed through an analysis filter bank followed by a particular decomposition level. At each decomposition level, the analysis filter bank consists of a high pass and low pass filter. The signal is passed through a series of such high pass and low pass filters. The output of high pass filter is known as detail coefficients (Cd) and contains the valuable information of the signal while as the output of the low pass filter is known as approximation filters (Ca) and contain least information of the signal. Detail coefficients are low scaled high frequency components while approximation coefficients are high scaled low frequency components. The frequency components that are not very prominent in the original signal will have very low amplitude and this part of the DWT signal can be discarded without loss of any valuable information, allowing data compression at higher data rates. During the decomposition procedure a signal is broken down into many lower resolution components [10-12]. Start Read a sound file Prediction Order LPC Encoding LPC Decoding Reconstruct Speech End Start Read a sound file Prediction Order DCT LPC Encoding LPC Decoding IDCT Reconstruct Speech End Start Read a sound file Prediction Order DWT ( db4 ) LPC Encoding LPC Decoding IDCT Reconstruct Speech End LPC LPC+DCT LPC+DWT Figure 1: Block diagram of the Proposed Hybrid Technique V. SIMULATION RESULTS The compression technique used in this paper has been verified on a number of test recorded female spoken speech signals (about 3-files). Table 1 shows performance measures of speech signals using different compression techniques including proposed hybrid techniques. To evaluate the performance of compression technique used, various performance parameters related a speech signal are calculated. RES Publication 2012 Page 18

4 These include % of zero coefficients, Retained Signal Energy, Mean Square Error, Peak Signal to Noise and Compression Ratio. The above quantities are calculated using the following formats: MSE = { }/N (10) where err is the error signal N is the size of original signal. PSNR = 10log10{max(A)/MSE} (11) where A is the original signal CR = x(n)/y(n) (12) Where x(n) is the length of original signal y(n) is the length of compressed signal. Figure 3: Comparative graph of the three techniques used for speech signals S2 Table 1: Performance of Speech Signals using different Compression Techniques Compression Technique Speech Signal S1 (3.12 sec long) Signal S2(1.77 seconds long) Signal S3(1.26 seconds long) LPC LPC + DCT LPC + DWT MSE X10-7 PSNR CR MSE x10-7 PSNR CR MSE x10-7 PSNR CR Figure 4: Comparative graph of the three techniques used for speech signals S3 Table 2: Comparison of PSNR and CR of proposed hybrid techniques with existing techniques Proposed Technique Technique used LPC LPC+DWT Mawla et al (2003) [3] Existing Techniques D. Ambika & V. Radha (2012) [2] DWT LPC DWT PSNR CR Figure 2: Comparative graph of the three techniques used for speech signals S1 Figure 5: Comparative performance of proposed LPC + DWT technique with existing DWT technique Figures 2-4 shows the comparison of various parameters (PSNR, CR, and MSE) between the three techniques used in our proposed work for. It can be clearly seen from the above bar graphs that the results obtained for LPC + DWT hybrid technique gives better performance in terms of PSNR and MSE than other two techniques. Moreover when compared to the other existing techniques as shown in Table 2 and Figure 5 above, it is found that the hybrid technique of LPC + DWT shows better results than the LPC and DWT when implemented individually. Thus it can be concluded that the hybrid technique LPC + DWT is best suitable for compression of speech signals with much high value of the peak signal to noise ratio (PSNR). The time domain waveform of the original and reconstructed signals using conventional and proposed techniques is shown in Figures 6-14, respectively. RES Publication 2012 Page 19

5 Figure 6: Original and reconstructed signal S1 using LPC Figure 11: Original and Reconstructed Signal S2 using LPC + DWT Figure 7: Original and reconstructed signal S1 using LPC + DCT Figure 12: Original and Reconstructed Signal S3 using LPC Figure 8: Original and reconstructed signal S1 using LPC + DWT Fig. 13: Original and Reconstructed Signal S3 using LPC + DCT Figure.9: Original and Reconstructed Signal S2 using LPC Figure 10: Original and Reconstructed Signal S2 using LPC + DCT Fig. 14: Original and Reconstructed Signal S3 using LPC + DWT VI. CONCLUSION Speech compression is one of the major contributions of information theory for efficient transmission of speech over computer and wireless networks. This work shows that wavelet decomposition in conjunction with other techniques such as LPC is promising compression techniques which make use of the elegant theory of wavelets. Several conclusions have been drawn from the observations: The two proposed techniques work better with clean source materials; noisy sound waveforms produce poor results and need additional processing techniques. RES Publication 2012 Page 20

6 Linear predictive technique causes time delay and some loss of quality. But they are negligible in terms of cost when compared with the advantages of storage space saving, smaller bandwidth requirement, lower power consumption and small product size. This proposed method could be classified in the field of symmetrical compression. This case occurs when the compression and decompression use basically the same algorithm but work in opposite directions. With these facts it can be concluded that the proposed scheme has a lot of scope in wireless communications where bandwidth and Quality of Service (QOS) are two important factors REFERENCES [1] Rabiner L.R, and Schafer R.W, Digital Processing Of Speech Signals. Prentice Hall. [2] Markos Zampoglou and Athanasios G. Malawas, Music Information retrieval in compressed audio files: a survey. New Review of Hymermedia and Multimedia Journal, IETE Technical Review. Volume 20, issue 3, pp Jul [3] M. V. Patil, Apoorva Gupta, Ankita Varma and Shikhar Salil, Audio and Speech Compression using DCT & DWT Techniques, International Journal of Innovative Research in Science, Engineering and Technology, Vol 2, ISSN: , Issue 5, May [4] Siva Nagu, T, K. Jyothi and V. Sailaja, Speech Compression for Better Audibility using Wavelet Transformation with Adaptive Kalman Filtering, Internatioanal Journal of Computer Applications , Vol 53-No.11, Sep [5] Harmanpreet Kaur and Ramanpreet Kaur, Speech Compression and Decompression using DWT and DCT, Int.J. Computer Technology & Applications, Vol 3, pp , Aug [6] Satish Kumar, O.P. Singh, G.R. Mishra, Saurab Kumar Mishra and Akanksha Trivedi, Speech Compression and Enhancement using Wavelet Coders,International Journal of Electronics Communication and Computer Engineering Volume 3, Issue 6, ISSN (Online): X, ISSN (Print): [7] Vidya Pawar and S. D. Apte, Piano note synthesis using Linear Prediction and Wavelet Transform International Journal of Digital Signal Processing ISSN [8] Khalid Sayood, Introduction to Data Compression. Morgan Kaufmann Publishers, Elsevier Inc., 2006 [9] Koul M., Linear Prediction Tutorial, Proceeding of IEEE, Vol.63, No.4, pp , Apr [10] Javaid A Sheikh, Shabir A Parah, Sakeena Akhtar, G.M. Bhat, Performance Evaluation and Comparison of Speech Compression using Linear Predictive Coding and Discrete Wavelet Transform, Commune 2015 International Conference on Advances in Computers, Communication and Electronic Engineering March 16-18, 2015, pp: , University of Kashmir. [11] Javaid A. Sheikh, Sakeena Akhtar, Shabir A. Parah, G. M. Bhat, On the Design and Performance Evaluation of Compressed Speech Transmission over Wireless Channel, 12th IEEE India International Conference (INDICON) on Electronics, Energy, Environment, Communication, Computers, Control (E3-C3), December, 2015, Jamia Millia Islamia, New Delhi. [12] Joebert S. Jacaba Audio Compression using modified Discrete Cosine Transform: the MP3 coding standard in Oct AUTHOR S BIOGRAPHY Javaid A. Sheikh has completed his M.Sc., M. Phil and Ph. D in Electronics from University of Kashmir, Srinagar in the year 2004, 2008 and 2012 respectively in the field of communications and Signal Processing. He is working as Assistant Professor in the department of Electronics and I. T University of Kashmir, Srinagar. His field of interest are Wireless Communications, design and development of efficient MIMO- OFDM based wireless communication techniques, Spread Spectrum modulation, Digital Signal Processing, Electromagnetics and Speech Processing and compression techniques. Sakeena Akhtar has completed her M.Sc and M.Phil in Electronics from University of Kashmir, Srinagar in the year 2012 and 2016 respectively in the field of Communication and Speech Processing and Speech Compression Techniques. She is presently pursuing her Ph.D. in the same field from the University of Kashmir. Shabir A. Parah has completed his M. Sc, M. Phil and Ph. D in Electronics from University of Kashmir, Srinagar in the year 2004, 2010 and 2013 respectively in the field of Signal processing and Data Hiding. He is working as Assistant Professor in the department of Electronics and I. T University of Kashmir, Srinagar. His field of interest are Signal Processing, Embedded Systems, Secure Communication and Digital design and Watermarking. He has published more than 60 research papers in International and National journals and conference proceedings. Prof. G. Mohiuddin Bhat obtained his M.Sc. (Electronics) from the University of Kashmir, Srinagar (India) in 1987, M.Tech. (Electronics) from Aligarh Muslim University (AMU), Aligarh (India) in 1993 and Ph.D. Electronics Engg. From AMU, Aligarh, (India) in He has served as Assistant Professor, Associate professor, Professor and Director in University Science Instrumentation Centre (USIC), University of Kashmir. Presently Prof. G. M. Bhat is Dean Faculty of Applied Science and Technology, University of Kashmir. He has worked in the area of Mobile Radio Communication, Spread Spectrum Communication and Neural Networks and has guided many research degrees leading to the award of M.Phil and Ph.D. RES Publication 2012 Page 21