Speech Compression Using Wavelets Team Members: Deepak Kumar, 3rd B-tech, E. C. E Branch, JNTU College of Engineering, Kakinada. E-Mail: [email protected] com Ch. Naresh Kumar, 3rd B-Tech, E. C. E Branch, JNTU College of Engineering, Kakinada. ABSTRACT Speech compression is the technology of converting human speech into an efficiently encoded representation that can later be decoded to produce a close approximation of the original signal. The wavelet transform of a signal decomposes the original signal into wavelets coefficients at different scales and positions.

These coefficients represent the signal in the wavelet domain and all data operations can be performed using just the corresponding wavelet coefficients. The major issues concerning the design of this Wavelet based speech coder are choosing optimal wavelets for speech signals, decomposition level in the DWT, thresholding criteria for coefficient truncation and efficient encoding of truncated coefficients. The performance of the wavelet compression scheme on both male and female spoken sentences is compared. On a male spoken sentence the scheme reaches a signal-to-noise ratio of 17. 5 db and a compression ratio of 3. 88, using a level dependent thresholding approach. 1. INTRODUCTION Speech is a very basic way for humans to convey information to one another. With a bandwidth of only 4 kHz, speech can convey information with the emotion of a human voice. People want to be able to hear someone’s voice from anywhere in the world as if the person was in the same room . As a result a greater emphasis is being placed on the design of new and efficient speech coders for voice communication and transmission.

Today applications of speech coding and compression have become very numerous. This paper looks at a new technique for analyzing and compressing speech signals using wavelets. Any signal can be represented by a set of scaled and translated versions of a basic function called the. mother wavelet. This set of wavelet functions forms the wavelet coefficients at different scales and positions and results from taking the wavelet transform of the original signal. Speech is a non-stationary random process due to the time varying nature of the human speech production system.

Non-stationary signals are characterized by numerous transitory drifts, trends and abrupt changes. The localization feature of wavelets, along with its time-frequency resolution properties makes them well suited for coding speech signals. 2. WAVELETS Vs FOURIER ANALYSIS A major draw back of Fourier analysis is that in transforming to the frequency domain, the time domain information is lost. The most important difference between these two kinds of transforms is that individual wavelet functions are localized in space.

In contrast Fourier sine and cosine functions are non-local and are active for all time t. 3. DISCRETE WAVELET TRANSFORM The Discrete Wavelet Transform (DWT) involves choosing scales and positions based on powers of two so called dyadic scales and positions. The mother wavelet is rescaled or. dilated, by powers of two and translated by integers. The numbers a(L, k) are known as the approximation coefficients at scale L, while d(j,k) are known as the detail coefficients at scale j. The approximation and detail coefficients can be expressed as: [pic] [pic] 3. 1. VANISHING MOMENTS

The number of vanishing moments of a wavelet indicates the smoothness of the wavelet function as well as the flatness of the frequency response of the wavelet filters (filters used to compute the DWT) Typically a wavelet with p vanishing moments satisfies the following equation, [pic] Wavelets with a high number of vanishing moments lead to a more compact signal representation and are hence useful in coding applications. However, in general, the length of the filters increases with the number of vanishing moments and the complexity of computing the DWT coefficients increases with the size of the wavelet filters. 3. 2.

FAST WAVELET TRANSFORM The Discrete Wavelet Transform (DWT) coefficients can be computed by using Mallet’s Fast Wavelet Transform algorithm. This algorithm is sometimes referred to as the two-channel sub-band coder and involves filtering the input signal based on the wavelet function used. 3. 2. IMPLEMENTATION USING FILTERS To explain the implementation of the Fast Wavelet Transform algorithm consider the following equations: [pic] The first equation is known as the twin-scale relation (or the dilation equation) and defines the scaling function. The next equation expresses the wavelet in terms of the scaling function.

The third equation is the condition required for the wavelet to be orthogonal to the scaling function and it translates the coefficients c(k) or {c0, .. , c2N-1} in the above equations represent the impulse response coefficients for a low pass filter of length 2N, with a sum of 1 and a norm of 1/squrerootof(2). The high pass filter is obtained from the low pass filter using the relationship [pic] where k varies over the range (1-(2N-1)) to 1. Starting with a discrete input signal vector s, the first stage of the FWT algorithm decomposes the signal into two sets of coefficients.

These are the approximation coefficients cA1 (low frequency information) and the detail coefficients cD1 (high frequency information), as shown in the figure below. [pic] The coefficient vectors are obtained by convolving s with the low-pass filter Lo_D for approximation and with the high-pass filter Hi_D for details. This filtering operation is then followed by dyadic decimation or down sampling by a factor of 2. Mathematically the two-channel filtering of the discrete signal s is represented by the expressions: [pic] These equations implement a convolution plus down sampling by a factor 2 and give the forward fast wavelet transform.

If the length of each filter is equal to 2N and the length of the original signal s is equal to n, then the corresponding lengths of the coefficients cA1 and cD1 are given by the formula: [pic] This shows that the total length of the wavelet coefficients is always slightly greater than the length of the original signal due to the filtering process used. 3. 3. MULTILEVEL DECOMPOSITION The decomposition process can be iterated, with successive approximations being decomposed in turn, so that one signal is broken down into many lower resolution components. This is called the wavelet decomposition tree.

The wavelet Decomposition of the signal s analyzed at level j has the following structure [cAj, cDj, cD1]. Looking at a signals wavelet decomposition tree can reveal valuable information. The diagram below shows the wavelet decomposition to level 3 of a sample signal S. [pic] Since the analysis process is iterative, in theory it can be continued indefinitely. In reality, the decomposition can only proceed until the vector consists of a single sample. Normally, however there is little or no advantage gained in decomposing a signal beyond a certain level.

The selection of the optimal decomposition level in the hierarchy depends on the nature of the signal being analyzed or some other suitable criterion, such as the low-pass filter cut-off. 3. 4. SIGNAL RECONSTRUCTION The original signal can be reconstructed or synthesized using the inverse discrete wavelet transform (IDWT). The synthesis starts with the approximation and detail coefficients cAj and cDj, and then reconstructs cAj-1 by up sampling and filtering with the reconstruction filters. [pic] The reconstruction filters are designed in such a way to cancel out the effects of aliasing introduced in the wavelet decomposition phase.

The reconstruction filters (Lo_R and Hi_R) together with the low and high pass decomposition filters, forms a system known as quadrature mirror filters (QMF). For a multilevel analysis, the reconstruction process can itself be iterated producing successive approximations at finer resolutions and finally synthesizing the original signal. 4. WAVELET SPEECH COMPRESSION The idea behind signal compression using wavelets is primarily linked to the relative scarceness of the wavelet domain representation for the signal. Wavelets concentrate speech information (energy and perception) into a few neighboring coefficients.

Therefore as a result of taking the wavelet transform of a signal, many coefficients will either be zero or have negligible magnitudes. Data compression is then achieved by treating small valued coefficients as insignificant data and thus discarding them. The process of compressing a speech signal using wavelets involves a number of different stages, each of which are discussed below. 4. 1. CHOICE OF WAVELET The choice of the mother-wavelet function used in designing high quality speech coders is of prime importance.

Choosing a wavelet that has compact support in both time and frequency in addition to a significant number of vanishing moments is essential for an optimum wavelet speech compressor. This is followed very closely by the Daubechies D20, D12, D10 or D8 wavelets, all concentrating more than 96% of the signal energy in the Level 1 approximation coefficients. Wavelets with more vanishing moments provide better reconstruction quality, as they introduce less distortion into the processed speech and concentrate more signal energy in a few neighboring coefficients. 4. 2.

WAVELET DECOMPOSITION Wavelets work by decomposing a signal into different resolutions or frequency bands, and this task is carried out by choosing the wavelet function and computing the Discrete Wavelet Transform (DWT). Signal compression is based on the concept that selecting a small number of approximation coefficients (at a suitably chosen level) and some of the detail coefficients can accurately represent regular signal components. Choosing a decomposition level for the DWT usually depends on the type of signal being analyzed or some other suitable criterion such as entropy.

For the processing of speech signals decomposition up to scale 5 is adequate, with no further advantage gained in processing beyond scale 5. 4. 3. TRUNCATION OF COEFFICIENTS After calculating the wavelet transform of the speech signal, compression involves truncating wavelet coefficients below a threshold This means that most of the speech energy is in the high-valued coefficients, which are few. Thus the small valued coefficients can be truncated or zeroed and then be used to reconstruct the signal. This compression scheme provided a segmental signal-to-noise ratio (SEGSNR) of 20 dB, with only 10% of the coefficients.

Two different approaches are available for calculating thresholds. The first, known as Global Thresholding involves taking the wavelet expansion of the signal and keeping the largest absolute value coefficients. In this case you can manually set a global threshold, a compression performance or a relative square norm recovery performance. Thus only a single parameter needs to be selected. The second approach known as By Level Thresholding consists of applying visually determined level dependent thresholds to each decomposition level in the wavelet transform. . 4. ENCODING COEFFICIENTS Signal compression is achieved by first truncating small-valued coefficients and then efficiently encoding them. One way of approach to compression is to encode consecutive zero valued coefficients, with two bytes. One byte to indicate a sequence of zeros in the wavelet transforms vector and the second byte representing the number of consecutive zeros. For further data compaction a suitable bit encoding format, can be used to quantize and transmit the data at low bit rates.

A low bit rate representation can be achieved by using an entropy coder like Huffman coding or arithmetic coding. 4. 5. DETECTING VOICED Vs UNVOICED SPEECH FRAMES In speech there are two major types of excitation, voiced and unvoiced. Voiced sounds are produced when air flows between the vocal cords and causes them to vibrate. Voiced speech tends to be periodic in nature. Examples of voiced sounds are English vowels, such as the /a/ in . bay and the /e/ in . see. Unvoiced sounds result from constricting the vocal tract at some point so that turbulence is produced by air flowing past the constriction.

Since unvoiced speech is due to turbulence, the speech is aperiodic and has a noise-like structure. Some examples of unvoiced English sounds are the /s/ in . so and the /h/ in . he.. In general at least 90% of the speech energy is always retained in the first N/2 transform coefficients, if the speech is a voiced frame However, for an unvoiced frame the energy is spread across several frequency bands and typically the first N/2 coefficients holds less than 40% of the total energy. Due to this wavelets are inefficient at coding unvoiced speech.

Unvoiced speech frames are infrequent. By detecting unvoiced speech frames and directly encoding them (perhaps using entropy coding), no unvoiced data is lost and the quality of the compressed speech will remain transparent. 4. 6. PERFORMANCE MEASURES A number of quantitative parameters can be used to evaluate the performance of the wavelet based speech coder, in terms of both reconstructed signal quality after decoding and compression scores. The following parameters are compared: 4. 6. 1. SIGNAL TO NOISE RATIO [pic] 4. 6. 2.

PEAK SIGNAL TO NOISE RATIO (PSNR) [pic] 4. 6. 3. NORMALISED ROOT MEAN SQUARE ERROR (NRMSE) [pic] 4. 6. 4. RETAINED SIGNAL ENERGY [pic] 4. 6. 5. COMPRESSION RATIOS [pic] 5. PERFORMANCE OF RECORDED SPEECH CODING A male and female spoken speech signals were decomposed at scale 3 and level dependent thresholds were applied using the Birge-Massart strategy. Since the speech files were of short duration, the entire signal was decomposed at once without framing. A summary of the performance is given below for the different wavelets used. [pic] [pic] CONCLUSION

Speech coding is currently an active topic for research in the areas of Very Large Scale Integrated (VLSI) circuit technologies and Digital Signal Processing (DSP). The Discrete Wavelet Transform performs very well in the compression of recorded speech signals. For real time speech processing however, its performance is not as good. Therefore for real time speech coding it is recommended to use a wavelet with a small number of vanishing moments at level 5 decomposition or less. The wavelet based compression designed reaches a signal to noise ratio of 17. 45 db at a compression ratio of 3. 88 using the Daubechies 10 wavelet.

The performance of the wavelet scheme in terms of compression scores and signal quality is comparable with other good techniques such as code excited linear predictive coding (CELP) for speech, with much less computational burden. In addition, using wavelets the compression ratio can be easily varied, while most other compression techniques have fixed compression ratios. References [1]. A. Chen, N. Shehad, A. Virani and E. Welsh, Discrete Wavelet Transform for Audio Compression, (current July. 16, 2001). [2]. Speech Compression Using Wavelets by Nikhil Rao [3]. S. Haykin, Communication Systems, John Wiley & Sons, New York, 2001.