A LPC-PEV Based VAD for Word Boundary Detection

Transcription

1 14 A LPC-PEV Based VAD for Word Boundary Detection Syed Abbas Ali (A), NajmiGhaniHaider (B) and Mahmood Khan Pathan (C) (A) Faculty of Computer &Information Systems Engineering, N.E.D University of Engg. & Tech., Karachi- Pakistan.. saaj@neduet.edu.pk (B) Faculty of Computer Science & Information Technology, N.E.D University of Engg. & Tech., Karachi- Pakistan. najmi@neduet.edu.pk (C) Faculty of Computer Science & Information Technology, N.E.D University of Engg. & Tech., Karachi- Pakistan. mkpathan@hotmail.com Abstract The speech is the prominent mode of communication and interaction among human being and machines. In most of the practical application of ASR, the input speech is contaminated by ambient noise. In this paper LPC-PEV algorithm is proposed for word boundary detection in the presence of noise. In the proposed algorithm, the variance of error is estimated by AR model using Yule Walker method. We used the speech and noise samples taken from TIDIGIT corpus and NOISE-92 databases respectively, and evaluated the performance of system using LPC-PEV and LPC residual error. The experimental results based on numerical values of the samples clearly evident that the boundary detection rate of proposed algorithm is far better than the LPC residual error in term of clean and noisy speech. Index Terms Predictive capability, Linear prediction coefficient, precision, Residual error, Prediction error variance (PEV). V I. INTRODUCTION AD is widely used to identify the presence of speech in an input signal bymarking the boundaries of speech and non-speech segments. Speech recognition systemcan be considerably improved by integrating a VAD module into the system. VAD is a classification problem in which features of the audio signal are used to separate the input speech and nonspeech. The end points accuracy is one of the major factors in recognition performance. Good end point detection leads to effective computation. Also, it leads to accurate recognition since proper end points will result in good alignment for model making and template comparison [1]. Aprecise VAD reduces the computation cost and response time of speech recognition systems and only detected speechframes are allowed to passinto the recognition algorithm. Conversely, accurate detection of the end points of the speech in the presence of ambient noise avoids the waste of speech recognition system evaluation on ensuing silence [2]. The noise reduction task can be improvedby removing the unwanted portion of the speech signal. Feature extraction is an essential VAD module andone of the dominant approaches of feature extraction is temporal domain. Thelinear prediction analysis is the best example of temporal domain.lpc based algorithm performs well in high-level Gaussian-like ambient noise, short pulses, transient pulses and low level noises. LPC parametersdescribe the spectral peaksby forming a perceptually attractive description of the spectral envelope. LP model models the human vocal tract as an infinite response that produces the speech signal. LPC are coefficients of an AR model of speech frame and the all pole representation of the vocal tract transfer function. LP is a well-known method for finding all-pole model parameters, LP spectral envelopes overemphasize and overestimate the high and low pitch (voiced) speech spectral power, thuscontaining unwanted sharp contours, and don t increase in spectral envelope modeling performance as order of the filter increased. LP model assumed that speech modeled as the output of an all pole filter and excitation to this filter is single impulse andrandom noise sequence. In real scenario these assumptions are not exactly valid for observed speech signal especially for voiced portion and not also valid for voicedspeech, where the excitation source is a pulse train of certain pitch period. As a result, the speech samples in the nearest region of pitch pulses are not anticipated well and the residual error is relatively high in the locality of the pitch pulse and LPC coefficients are estimated by autocorrelation or covariance methods contain a certain amount of error when compared to actual value. This error is called prediction error. Exploiting this error we propose a statistical approach which is used to monitor the variance of the prediction error. The proposed LPC-PEV measures the prediction capability of LP model and measures the precision of LP model predictions, which prevents overestimates and overemphasizes issues in speech spectral power due to high pitch voiced and medium.this precision value will assist classifier to efficiently improve the boundary detection rate of the spoken word in the presence of noise.rest of the paper is organized as follows. In section II, explain the LPC coefficients computation used for LPC analysis using autocorrelation methods. Proposed LPC-PEV algorithm framework is presented in section III. In section IV, we discuss the VAD module using proposed LPC-PEV algorithm. We present our experimental results and discussion in section V. Finally, the conclusion is drawn in section VI.

2 15 II. LPCAND RESIDUAL ERROR ANALYSIS LPC family among the speech recognition approaches is well known for its performance and effective way of estimating the main parameters of speech signal. The parameters estimation method of an autoregressive model for speech is well established in [3]. For our discussion,consider a voiced speech signal {S (n) = s 1, s 2.s N } having N samples which can be modeled as a summation of periodic signals. S () = c Cos(ω k + ) (1) Where ω the pitch frequency of the signal is, c are the amplitudes at the harmonics and L is the no. of harmonics. L=, where f is typically 8 khz for speech. In LPC analysis work under the assumption that the current sample is predicted approximately by a linear combination of p past samples, S () = a S () (2) Where a 1, a 2 a p are the predictor coefficients, p is the order of the LPC analysis which is assumed to be 10 for speech samples at 8 khz. The predictor coefficients a k are calculated by reducing the sum of squared differences over the finite interval between the real speech samples and the linearly predicted ones. The error between the real and the predicted sample value is called the residual error and is given by e S () S () e = S () ( a S () ) (3) The power spectrum envelope will be approximately flat, due to low value of short-term correlation between samples of residual signal envelope of its power spectrum will be approximately flat. By taking Ztransform of (3), E(z)= A(z)S(z) (4) A(z)=1+ a Z (5) Where A(z) is the output of an LP FIR filter which eliminates the short-term correlation present in the speech signal which levels the spectrum. E(z) and S(Z) are the residual signal and speech signals in z-domain respectively. It is actually the error signal between the current sample values of the input and a predicted value based on p past sample values. E(z) has an approximately flat spectrum, LPC analysis modeled the short-time power-spectral envelope of the speech signal using an autoregressive model. H(z) = 1/A(z) (6) The LPC coefficients are computed from the speech sample and can beobtained by reducing the total-squared residual error. = (7) The summation range of (7) based on the two methods namely, autocorrelation and covariance,which are used for LPC analysis. We used the autocorrelation method in our experiments. In the autocorrelation method, short term LP analysis can be achieved using windowing the speech signal. In this analysis, the hamming window function is chosen over rectangular window function by assuming the samples outside this window consider being zero. In (7) error minimization criterion leads us to following equations: = 1 i p, (8) = (9) is the window function, N is thesamples duration and(8) called the Yule-Walker equation [4]. The matrix forms of the equations are: = (10) (11) =[,,., ] (12) =[,,., ] (13) Autocorrelation matrix in (11) based on toeplitz structure facilitates the solution of Yule Walker equation in (8) and (10) for LP coefficient by using Levinson-Durbin [5] and Schur[6] algorithms. Theeplitz structure also assure that the poles of the LP synthesis filterh(z) resulting from autocorrelation method should be inside the unit circle to maintain the stability of the LP synthesis filterh(z). III. VARIANCE OF THE PREDICTION ERROR LPC analysis works under the assumption that speech signal can be modeled as theoutput of an all-pole filter H(z) in (6).Excitation to this filter value is assumed to be either voiced speech or unvoiced speech (for single impulse or a white random noise respectively).in a real practice above mentioned assumptions are not exactly valid for observed speech signal. As a result LP coefficient projected by means of autocorrelation method contains a certain amount of error which is called Prediction error. Due to this prediction error near pitch pulses the speech samples are not predicted well and the prediction error is relatively large in the region of these pitch pulses [7], which affects the estimation accuracy of LPC analysis [8]. The Levinson-Durbin algorithm can be used to compute the filter coefficients a k of the LP method that minimizes the MSE. For voiced signal, the LP spectrum () can be observed as a spectral envelope whose samples provide an estimation of the voiced speech power [3]. () = (14) P EV is the variance of the prediction error. From the speech model discussed in (1), the speech signal has a correlation sequence; U ()= ( ) (15) The power spectrum shows a discrete line spectrum with the powers at the exponential frequenciesω k, i.e. ()= [(+ )+( )] (16) The variance of prediction error P EV is the Mean Squared Error (MSE) of the output of the filter A(z) and is given as; = ( ) (ω)dω = A(e ) (17) In general, variance of the prediction error P EV is a very

3 International Journal of Electrical & Computer Sciences IJECS IJECS-IJENS IJENS Vol: 12 No: 02 0 useful way to examine the predictive capability of a model. It provides a degree of the precision of a model's predictions.lp is a well-known known method for obtaining all all-pole model parameter. LP spectral envelopes over emphasize and overestimate the high and medium pitch voiced speech spectral power, thus featuring the unwanted sharp contours the spectral envelope modeling performance decreases as the order of filter increases[9]. IV. LPC-PEV BASED VAD MODULE 16 points detection algorithms are widely used as preprocessing of the sound waves in order to cut off the unwanted portion of speech present at the beginning and end of the speech sample. The VAD decision is normally based on feature vectors. For VAD decision, ision, we assume that the speech and noise are additive in nature. VAD normally consists of the following modules: Feature extraction, Start and end point decisions and decision smoothening. In our experiment, algorithm that is used in VAD to detect the end nd points of spoken word is LPCLPC PEV. VAD module includes end points detection of speech. E End Fig.11 Flow Diagram of LPC-PEV based VAD module Fig.1provides the overview of the LPC-PEV PEV based VAD module. Pre-emphasis emphasis is done to flat the signal spectrally. A first order high pass FIR filter is used for emphasis of the higher frequency components. For this purpose, a single coefficient digital filter with coefficient values is used. Spectral normalization is performed to compensate the distortion effect in the speech signal generated by linear convolution, which may be a result of filtering by the human vocal tract, room acoustics and the transfer fu function of aninputdevice like microphone. The spectral subtraction is done to cancel out any stationary noise in the input speech signal. It is a commonly used technique, and hence it is taken from the Voice box toolbox for MatLab. Framing is done to decompose pose the speech signal into a series of overlapping frames. Both the algorithms are designed for a frame length of 10 milliseconds. The sampling frequency is 8 KHz, therefore, there will be 10m*8k = 80 samples per frame. Also, it is better to use overlapping ng windows to ensure better temporal continuity in the transform domain; therefore, 75% overlapping rate is used. The Auto Regression model parameter was initially selected as 10 and a smoothening order as 20. We then estimated the AR all pole model using YuleWalker method [10]. The end point decision module (start and end) defines the rules for assigning speech or silence class to the LPC-PEV PEV features extracted in the previous block. The start and endpoint decisions are based on thresholds which are different erent for each speech sample. In order to smooth the decision curve, a moving average is computed of the LPCLPC PEV values. This is taken as the final decision curve. The decision smoothening is done to make the algorithm robust against noise. Also, some hang-over over algorithms use this smoothed decision curve in order to recover speech that is masked by the ambient noise [11]. V. EXPERIMENTAL RESULTS Experiments were done for word boundary detection in the presence of noise using two feature extraction techniques; LPC residual error (en)and the proposed LPC-PEV LPC (PEV). The comparison is done using MatLab tool version 10.0.MatLab performs computationally expensive tasks much faster than the other Conventional programming languages (e.g C and C++).We used speech processing ssing toolbox (Voice ( box tool) consisting of various MatLab routines. The speech database used in the experiments contains the isolated digits in English language taken from TIDIGIT Corpus [12]. The TIDIGITS corpus consists of more than 25 IJENS

4 17 thousand digit sequences spoken by over 300 men, women, and children. The data was collected in a quiet studio environment and digitized at 20 khz. In our experiment,we used 10 utterances of each digit from 0-9 in an approximately clean environment with a sampling frequency of 8 khz. Various types of noises such as babble, white noise, pink noise and Volvo (car) were collected from NOISEX-92 noise-inspeech database [13,14]. Two algorithms were developed in MatLab for comparison; one extracts the LPC-PEV from the input speech signal, whereas the other gives its boundary decisions on the basis of Residual error. The algorithm uses frame length of 10 milliseconds and 75% overlapping rate. The auto regression model parameters was initially selected as 10 and a smothering order 20, then we estimated the AR all pole model using Yule-Walker method. In this work, we have investigated both VAD based algorithms with clean isolated digit from 0-9 and isolated digit(0-9) in the presence of babble noise, white noise, pink noise and car (Volvo) noise.in this paper, we selected clean isolated digit zero and the same digit with white noise for graphical representation. In Fig.2 and Fig.3, graphical results represent the detected word based on LPC residual error algorithm with clean isolated digit zero and isolated digit zero in the presence of white noise respectively. Similarly in Fig.4 and Fig.5 represent the detected word based on LPC-PEV algorithm with clean isolated digit zero and isolated digit zero in the presence of white noise respectively. The red vertical lines show the start and end point positions in terms of samples of detected isolated digit. The waveforms are of the isolated digit 0. Duration of the speech is seconds. The curve with varying values show the LPC-PEV and the LPC residual error in their respective graphs, whereas the smoothed curve shows the decision curve on which the boundary decisions are made. The bottom green horizontal line indicatesthe lower threshold whereas the upper blue horizontal line indicates the upper threshold. Fig.3 LPC residual error algorithm (Digit zero with white noise) Fig.4 LPC-PEV algorithm (clean isolated digit zero ) Fig.5 LPC PEV algorithm (Digit zero with white noise) Fig.2 LPC residual error algorithm (clean isolated digit zero ) From the above figures we can make the comparison of endpoint detection based on LPC residual error and LPC-PEV, when the isolated digit is clean and in the presence of white noise. It can be seen that the detected boundaries of the word

5 18 segment using LPC-PEV indicates very successful endpoint detection in case of both clean and noisy speech as compared to the other technique. Table.1 to Table.10 provides the comparative analysis of word boundary decision from digit 0 to digit 9 respectively. To develop this comparison of boundary decision, we used start and end boundary points where the speech signal actually begins and ends. The tables describe the numerical value of the start and end point of the isolated digit (0-9) in terms of sample. The boundary points in the presence of a number of background noises such as babble, car, white and pink are also observed. Table.1 Boundary Decision for digit zero Table.2 Boundary Decision for digit one Table.3 Boundary Decision for digit two

6 19 Table.4 Boundary Decision for digit three Table.5 Boundary Decision for digit four Table.6 Boundary Decision for digit five

7 20 Table.7 Boundary Decision for digit six Table.8 Boundary Decision for digit seven

8 21 Table.9 Boundary Decision for digit eight Table.10 Boundary Decision for digit nine The comparison has been made on the basis of thenumeric values of starting and ending point of the samples. The starting point should be closest to the spoken word sample and end point should be away from spoken word sample for avoiding the speech loss and accurate word detection. Our experimental results show that the numerical values of the starting and ending point closest to and away from the spoken word respectively. From the comparative analysis of the boundary decision for isolated digit clearly evident that the boundary detection rate of LPC-PEV is far better than LPC Residual error both in term of clean and noisy speech. The study presented in this paper based on the preconditioned speech and noise data bases. This reflects the limitation of the LPC-PEV based VAD, when the speech and noise are encounter in the real world environment. VI. CONCLUSION In this paper, we presented a LPC-PEV based VAD algorithm for word boundary detection in the presence of noise. LPC coefficients are estimated by autocorrelation method contain a certain amount of error when compared to actual value. This error is called prediction error. Exploiting this error we propose a statistical approach which is used to monitor the variance of the prediction error. The proposed LPC-PEV measures the prediction capability of LP model and measures the precision of LP model predictions, which prevents overestimates and overemphasizes issues in speech spectral power due to high pitch voiced and medium. Experiments have been performed with TIDIGIT corpus and NOISEX-92 speech and noise-in speech databases respectively. The performance of the system is evaluated using LPC-PEV and LPC Residual error based VAD algorithm. From the numeric values of starting and ending point of the samples, experimental results indicate that the boundary detection rate of proposed LPC-PEV is far better than LPC residual error in term of clean and noisy speech. From future research aspect, it will also be

9 22 interesting to study the LPC-PEV based VAD to improve the robustness of the speech recognition system and combine this feature with any margin based learning algorithm to improve the generalization capability of the acoustic model. REFERENCES [1] LingyunGu, Stephen A. Zahorian, A New Robust Algorithm for Isolated Word detection, Proc. ICASSP, [2] M. Karnjanadecha, Stephen A. Zoahorian, Signal Modeling for Isolated Word Recognition, Proc. ICASSP,pp ,1999. [3] J. Makhoul, Linear prediction: A tutorial review," Proc. of the IEEE, vol. 63, no. 4, pp , [4] S.M. Kay,Modern Spectral Estimation: Theory and Application. Englewood Cliffs, NJ: Prentice Hall, [5] L. R. Rabiner and R. W. Schafer, Digital Processing of Speech Signals. Englewood Cliffs, NJ:Prentice-Hall, [6] J. Schur, UberPotenzreihen, die in Innern des Einheitskreisesbeschranktsind," J. fuer diereine and AngewandteMathematiek, vol. 147, pp , [7] K.K.Paliwal, W.B. Klejin, Quantization of LPC parameters. [8] B.S. Atal, Linear prediction of speech - Recent advances with applications to speech analysis,"in Speech Recognition, D.R. Reddy, Ed. New York: Academic Press, pp , [9] Manohar. N,Murthi and B. D. Rao, All-Pole Modeling of Speech Based on the Minimum Variance Distortion less Response Spectrum, IEEE TransactiononSpeechand Audio Processing, vol. 8, no. 3, May [10] S.L. Marple, Digital Spectral Analysis with Applications. Englewood Cliffs, NJ: Prentice Hall, [11] L. Karray, A.Martin, Toward improving speech detection robustness for speech recognition in adverse environments, Speech Communication, no. 3, pp [12] [13] Spib.rice.edu/spib/select_noise. [14] A.P. Varga, H.J.M Steeneken, M. Tomlinson, D.Jones, "The NOISEX-92 Study on the Effect of Additive Noise on Automatic Speech Recognition", In Technical Report, DRA Speech Research Unit,1992.