AUDIO ENCODING USING HUANG AND HILBERT TRANSFORMS K. Khaldi 1, 2, 4 , A.O. Boudraa 2, B. Torr sani 3, Th.

Chonavel 4 and M. Turki 1 e Unit Signaux et Syst` mes, ENIT, BP 37, Le Belvedre 1002 Tunis, Tunisia. e e IRENav (EA 3634), Ecole Navale, BCRM Brest, CC 600, 29240 Brest Cedex 9, France. 3 Universit de Provence, LATP, CMI, 39 rue F. Joliot-Curie, 13453 Marseille Cedex 13, France. e 4 Institut T l com; T l com Bretagne, LabSTICC UMR, BP 832, 29285 Brest Cedex, France. ee ee
2 1

(Kais.Khaldi,thierry.chonavel)@telecom-bretagne.eu, [email protected] [email protected], [email protected]

ABSTRACT In this paper an audio coding scheme based on the Empirical Mode Decomposition (EMD) in association with the Hilbert transform is presented. The audio signal is decomposed adaptively into intrinsic oscillatory components by EMD called Intrinsic Mode Functions (IMFs) and the associated instantaneous amplitudes and the instantaneous phases calculated. The basic principle of the proposed approach consists in encoding the instantaneous amplitudes and the instantaneous phases. The decoder recovers the original signal after IMFs reconstruction by demodulation, and their summation. The compression method is applied to different audio signals, and results compared to MP3 and to wavelet approaches.

2. HUANG TRANSFORM: EMD The EMD breaks down any signal x(t) into a series of IMFs through an iterative process called sifting; each one with a distinct time scale [5]. The decomposition is based on the local time scale of x(t), and yields adaptive basis functions. The EMD can be seen as a type of wavelet decomposition whose subbands are built up as needful to separate the different components of x(t). Each IMF replaces the signals detail, at a certain scale or frequency band. The EMD picks out the highest frequency oscillation that remains in x(t). By denition, an IMF satises two conditions : 1. Number of extrema and the number of zeros crossings may differ by no more than one. 2. Average value of the envelope dened by local maxima, and the envelope dened by local minima, is zero. Thus, locally, each IMF contains lower frequency oscillations than the just extracted one. The EMD does not use a predetermined lter or a wavelet function, and is a fully datadriven method [5]. To be successfully decomposed into IMFs, the signal x(t) must have at least two extrema (one minimum and one maximum). The IMFs are obtained using the following algorithm (sifting process) [5]: identify all extrema of x(t). interpolate between minima (resp. maxima), ending up with some envelope e min (t) (resp emax (t)). compute the average m(t) = (e min (t) + emax (t))/2. extract the detail d(t) = x(t) m(t). iterate on the residual m(t). Signal d(t) is a true IMF, if it satises conditions (1) and (2). 3. ANALYTIC SIGNAL With the Hilbert transform, H[.], the analytic signal z(t) corresponding to s(t) is given by : z(t) = s(t) + iH[s(t)] = a(t)ei(t) (1)

1. INTRODUCTION Audio signal compression of hight quality, and at low bit rate has become very important in many applications, such as digital audio broadcasting, multimedia and satellite TV, that request a lower bit rates and high delity. Different coding methods has been proposed for reducing the bit rate [1]-[2]. Furthermore, new methods of audio compression based on wavelet have been proposed in to reduce bit rate requirements [3]-[4]. However, a limit of the wavelet approach, is that the basis functions are xed, and thus do not necessarily match all real signals. In this work we investigates the interest of the EMD or Huang transform for audio encoding. The EMD has been introduced by Huang et al. [5] for analyzing data from non-stationary and nonlinear processes. The major advantage of the EMD is that the basis functions are derived from the signal itself. Hence, the analysis is adaptive in contrast to the traditional methods where the basis functions are xed. The basic idea of the proposed method is to encode the instantaneous amplitude (IA) and the instantaneous phase (IP) for each IMF, exploiting the smoothness of these instantaneous quantities. This method is applied to audio signals, and the results are compared to the wavelet and MP3 approaches.

where the given time series s(t) is the real part of (1), and the imaginary part is the Hilbert transform of s(t), H[s(t)] = 1 PV
+

4.2. HHT After adaptive segmentation, each audio frame x(t) is decomposed into sum of IMFs by the EMD, as follows:
L

s( ) d t

(2) x(t) =

IMFj (t) + rL (t)

(5)

PV is the Cauchy principal value of the integral. An analytic signal represents a rotation in the complex plane with the radius of rotation a(t) and the IP (t), where a(t) = [s(t)]2 + H[s(t)]2 and (t) = tan1
H[s(t)] s(t)

The IA a(t) and IP (t) of IMFs of audio signal are slowly varying, which is not true for general audio signals. The combination of the EMD applied to s(t) to generate IMFs, and the Hilbert transform of each IMF is called the Hilbert-Huang Transform (HHT).

where IMFj (t) is the j th IMF and rL (t) is the residual. The L value is determined automatically using standard deviation SD as stopping criterion which usually is set between 0.2 and 0.3 [5]. An example of decomposition of an audio frame is illustrated in gure 2. For each IMF, IP (n) and the IA a(n), are determined using Hilbert transform. Figure 3 shows the IA and IP of an IMF. 4.3. Encoding For class of audio signals studied, it is found that IA of IMFs are correlated. An example of such correlations is shown in gure 4. So, AR model is used to efciently exploit this temporally correlated information.
p

4. PROPOSED APPROACH Compared to our recently published works [6],[7], where essentially extrema are encoded, in the present work IA and IP which are valuable pieces of information are exploited for coding. 4.1. Segmentation In the proposed method, the audio signal is rst segmented adaptively into frames where it remains quasi stationary within each frame. This segmentation is based on the Local Entropic Criterion (CEL) which is a non parametric detector. The CEL at instant n for a signal x(n) is given by [8]: CELx (n) = Exc (n) [Exl (n) + Exr (n)] |Exc (n)| (3)

a(n) =
k=1

c(k)a(n k) + (n)

(6)

where Exc (n), Exl (n) and Exr (n) denotes the entropies of the principal window and the left and right sub-windows respectively. Exc (n) = Ex[n N ,n+ N 1] , 2 2 Exl (n) = Ex[n N ,n1] , 2 Exr (n) = Ex[n,n+ N 1] .
2

where [1, c(2), ..., c(p)] are the coefcients of the model and (n) is stationary zero mean input sequence that is independent of past outputs. Analysis of variation of IMFs IP show that for coding the classical scalar quantization can be used. Thus, only extrema of IP are encoded. Figure 5 shows the extrema (red circles) of an IMF that are coded. This information corresponds to encoding zero crossings of the imaginary parts of IMFs. Finally, the encoded coefcients of the IA and extrema of the IP is improved, by using lossless compression such as Huffman or Lempel-Ziv encoding techniques to store data. These techniques account for probability of occurrence of encoded data to reduce the number of bits allocated to. Although Lempel-Ziv is not optimum, the decoder does not require the encoding dictionary [9]. 4.4. Decoding The decoder requires only the encoded extrema of IP and calculates the remaining phase values by linear interpolation. IA is also generated by linear prediction. Finally, the estimated IMF is calculated as follows: IMF(n) = |(n)| cos((n)) a (7)

Shannon entropy of a signal x(n) in the interval [0, N 1], Ex[0,N 1] , is dened by :
N 1

Ex[0,N 1] =
k=0

|X(k)|2 log |X(k)|2

(4)

The audio frame is constructed by IMFs summation and the decoded audio signal is obtained by frames concatenation. 5. RESULTS The method is tested on different audio signals, sampled at 44.1 Khz. The results are compared to the MP3 et wavelet

where X(k) is the discrete Fourier transform of x(n). So the CEL has a value in the range of -1 to 1. A transient in the signal is characterized by a CEL > 0. An example of CEL variations for an audio frame guitar is shown in gure 1.

approaches. As criteria to evaluate the performance of the method, Signal to Noise Ratio (SNR) and Compression Ratio (CR), Subjective Difference Grade (SDG) and instantaneous Perceptual Similarity Measure (PSMt) are used [10]. Due to its good behavior for audio encoding, compared to other wavelets, the Daubechies wavelet of order 8 is used [4]. Table 1 shows the variation of TC and SDG against the number of AR level. So, it is clear that order 9 represents a good compromise between the TC and listening quality (SDG). Table 1. Variations of the TC and the SDG over the AR order. order 5 7 9 11 13 15 guitar TC SDG 13.40:1 -2.17 11.39:1 -1.08 10.15:1 -0.85 8.94:1 -0.83 8.14:1 -0.71 7.51:1 -0.63 violin TC SDG 13.35:1 -2.87 11.72:1 -1.91 9.96:1 -1.09 8.70:1 -1.05 7.74:1 -1.01 7.01:1 -0.92 sing TC 16.43:1 13.20:1 11.30:1 9.48:1 8.31:1 7.39:1 SDG -2.32 -1.12 -0.75 -0.73 -0.67 -0.51

6. CONCLUSION In this paper, a new coding method combining Huang and Hilbert transforms is presented. The estimated IP and IA of the extracted IMFs are used for audio signals coding. Obtained results for different audio signals show that the proposed method, performs better than the wavelet and MP3 approaches, and conrm our previous ndings [6],[7]. These results also show the interest of the EMD as basis for signals coding. To conrm the obtained results and the effectiveness of the EMD-compression approach, the scheme must be evaluated with a large class of audio signals and in different experimental conditions, such as sampling rates, sample sizes. 7. REFERENCES [1] J.D. Johnston, Transform coding of audio signals using perceptual criteria, IEEE. Select Areas Commun., vol. 6, pp. 314323, 1988. [2] P. Noll, MPEG digital audio coding, IEEE Sig. Process. Magazine, vol. 14, no. 5, pp. 5981, 1997. [3] P. Srinivasan and L.H. Jamieson, High quality audio compression using an adaptive wavelet packet decomposition and psychoacoustic modeling, IEEE Trans. Sig. Process., vol. 46, no. 4, 1998. [4] P.R. Deshmukh, Multiwavelet decomposition for audio coding, IE(I) Journal-ET, vol. 87, pp. 3841, 2006. [5] N.E. Huang et al., The empirical mode decomposition and Hilbert spectrum for nonlinear and non-stationary time series analysis, Proc. Royal Society, vol. 454, no. 1971, pp. 903995, 1998. [6] K. Khaldi, A.O. Boudraa, M. Turki, T. Chonavel and I. Samaali, Audio encoding based on the empirical mode decomposition, in EUSIPCO, Glasgow, 2009. [7] K. Khaldi, A.O. Boudraa, M. Turki and T. Chonavel, Codage audio perceptuel a bas d bit par d composition ` e e en modes empiriques, in GRETSI, Dijon, France, 2009. [8] G. Gonon, S. Montr sor and M. Baudry, Segmentation e multibande adapt e bas e sur le critre entropique local e e pour le codage audio, in GRETSI, Toulouse, 2001. [9] T. Welch, A technique for high-performance data compression, Computer, vol. 17, pp. 819, 1984. [10] R. Huber and B. Kollmeier, PEMO-Q a new method for objective audio quality assessment using a model of auditory perception, IEEE Trans. Audio, Speech and Language Process., vol. 14, no. 6, 2006.

Table 2, shows that the improvement in TC provided by the proposed method varies from 9.96:1 to 11.3:1 than the TC achieved by wavelets and MP3. Even for a sing signal, we still can observe the effectiveness of the proposed method in compression. A careful examination of the results reported in Table 2, shows that the proposed approach performs remarkably better than wavelet and MP3 methods. Furthermore, when listening the decoded signal, the proposed method produces lower noise compared to the wavelet method and MP3. This result is shown in table 2, when we see the acquired SDG values depending to TC is better than the other methods. The obtained results show the interest to encode both IA and IP. Table 2. Compression results of audio signals (guitar, violin and sing) by the proposed approach, MP3 and the wavelet. Signal Cr SNR SDG PSMt Cr SNR SDG PSMt Cr SNR SDG PSMt guitar 10.15:1 20.27 -0.85 0.89 9.42:1 20.17 -1.51 0.85 7.37:1 21.84 -0.79 0.92 violin 9.96:1 20.41 -1.09 0.84 9.83:1 19.65 -1.76 0.83 7.84:1 19.72 -1.05 0.86 sing 11.3:1 22.86 -0.75 0.91 10.11:1 23.43 -1.94 0.81 6.92:1 23.69 -0.67 0.96

MP3

Wavelet

EMD

4 Value of Inst phase

2 0 2 4

value of CEL

0.2 0

100

200

300

400

500

600

700

0.4
0.2 0.4 0.6 0.8 1

Value of Inst amplitude

0.3 0.2 0.1 0

100

200

300 Time

400

500

600

700

Fig. 1. CEL variation of the audio frame guitar.

Fig. 3. Instantaneous phase and instantaneous amplitude of IMF3.

1 0.9

Signal
Value of autocorrelation coefficient

1 0 1 0.1 IMF1 0 0.1 0.2 IMF2 0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

0 0.2 0.5 0 100 200 300 400 500 600 700

100

200

300 Time

400

500

600

700

IMF3

0 0.5 0.5 0 100 200 300 400 500 600 700

Fig. 4. Autocorrelation function of instantaneous amplitude of IMF3.

IMF4

0 0.5 0.1 0 100 200 300 400 500 600 700

4 3 2 Value of Inst phase 1 0 1 2

IMF5

0 0.1 0 100 200 300 400 500 600 700

Residual

0.1 0 0.1 0 100 200 300 Time 400 500 600 700

3 4

Fig. 2. Decomposition of an audio frame by EMD.

100

200

300 Time

400

500

600

700

Fig. 5. Instantaneous phase and their extrema of the IMF3.