International Journal of Scientific & Engineering Research, Volume 6, Issue 4, April-2015 1

ISSN 2229-5518

Effect of Pre-processing Stages on Recognition

Accuracy of a Developed Isolated Word

Recognition System

Animasahun, I, O*. and Popoola, J.J.

Department of Electrical and Electronics Engineering, Federal University of Technology Akure, Ondo State, Nigeria E-mail:
ianimasahun8619@gmail.com; jidejulius2001@gmail.com

*Corresponding Author

AbstractThe major observed cause of inaccuracy of the earlier developed word recognition systems is the lack of sufficient combination of pre- processing stages adopted before extracting the speech features. In investigating this observed deficiency, two isolated word recognition systems were developed. The first was developed without pre-processing stages while the second was developed with pre-processing stages. The paper combine d analog to digital conversion, end point detection and template preparation as paramount pre-processing techniques in developing the second word recognition system. The analog to digital conversion was used to obtain the digitized speech using a Gold wave software while linear pattern classifier was used as the end point detection technique to remove silence at the beginning and the ending of the digitized speech. Reliable master templates of the three speech samples words: count, stop and down were prepared using average template method. The performance evaluation carried out on the developed word recognition systems show a variation on the recognition accuracies of the three words employed with an average recognition of 23.3% and 76.7% without and with pre-processing stages respectively. The result of the study shows that pre-processing stages have significant effect on the accuracy of the word recognition system. Also, the results from the study show that differences in both ascent and pitch of the speakers have effects on the performance of the developed word recognition system.

Index TermsAverage Template Method, Dynamic Time Warping (DTW), Mel Frequency Cepstral Coefficient (MFCC), Nearest Neighbour Classi- fication, Probability Density Function, Speech recognition, Nearest Neigbour Classiication (NNC).

.

1 INTRODUCTION

—————————— ——————————
peech recognition has been an area of research over dec- ades. There are three types of speech recognition systems namely isolated, connected and complex word recognition systems. This paper focusses on the development of an isolat- ed word recognition system. Isolated or discrete word recogni- tion systems are systems that recognize digit words. A lot of works have been done which has to do with the development of isolated speech recognition system [1-3]. The targets of any developed recognition system are high recognition accuracy and run time. There are several reasons for the reduction in the recognition accuracies of most developed speech recogni- tion systems [4]. The major cause of inaccuracy is the lack of sufficient combination of pre-processing stages adopted before extracting the speech features. Another reason for the defi- ciency in the recognition accuracy of some of the earlier devel- oped speech recognition systems is inefficient pre-processing algorithms. As reported in [4] and [5], some of the crucial pre- processing stages needed to extract features necessary for speech recognition are analogue to digital conversion, silence removal/end point detection and the template preparation
stages.
In [1], speech activated appliance was developed without car-
rying out end point detection and also vector quantization
approach was used in template preparation. The recognition accuracies of the words used are low due to inefficient tem- plate preparation technique and the avoidance of end point detection technique.
This shows that before reasonable analysis can be carried out on speech, it must first be acquire in digital form as well as being recorded in a low noise environment [6]. Digital record- ers such as Audacity, Adobe Audition, and Goldwave can be used. Audacity has editing properties but has fixed sampling frequency and the digitized speech cannot be read into Matlab file. Adobe Audition also has the limitation of fixed sampling frequency. Goldwave on the other hand, is preferable, since it allows the user to fix the sampling frequency and also has ed- iting properties to enhance the quality of the digitized speech. One of such editing properties is the removal of background noise and the unvoiced segment of the digitized speech sam- ple, which is a fundamental step in speech recognition systems [4].
In addition, there are different approaches that have been de- veloped for detecting the speech end points in speaker ’s utter- ance. Two of them are energy based method which uses zero crossing rate and short time energy functions. The limitation

IJSER © 2015 http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 4, April-2015 2

ISSN 2229-5518

in the energy based method is that there are still errors in- curred in finding the exact end point of utterance [7]. Abso- lute Energy and Teager Energy (AETE) algorithm was pro- posed by Stephen [8], which is still far from perfect because of effect of background noise, which sometimes cause the algo- rithm to make the wrong start point sooner than the true start- ing point. A newly developed algorithm that adopted uni- dimensional Mahalanobis Distance and uses statistical proper- ties of background noise as well as physiological aspect of speech production was formulated in [9].
The second activity to be carried out in order to enhance accu- rate recognition system is template preparation. One of the main problems in speech recognition systems is the prepara- tion of reliable reference templates for the set of words to be recognized [5]. Vector quantization (VQ) has been widely used as a solution to prepare reliable templates for the Dynamic Time Warping (DTW) based speech recognition systems but it requires many training examples to prepare a reliable code- book. In order to enhance the computational efficiency, a sim- ple method is to use single reference template per word [5]. Another approach is by using average template method which according to [5] helps to mitigate bad samples from good samples by using a master template for each speech sample prepared from several speech samples. The paper focuses on the pre-processing algorithms and discussed the results of each algorithm. Also, pitch detection algorithm was also im- plemented in this paper in order to justify that the fundamen- tal frequencies of each speaker varies and that it has a way of affecting the speech recognition accuracy. The speech features after these pre-processing stages were extracted and the speech samples classification was achieved using decision log- ic.
The rest of the paper is organized as follows: the pre- processing algorithms are presented in section 2 while the de- cision logic approach used during the testing phase are pre- sented in section 3. The result and discussion are presented in section 4 while concluding remarks are presented in section 5.

2 Pre-Processing Stages for Development of

Speech Recognition

The pre-processing stages and the methods of implementation are presented in this section. The pre-processing stages discussed in the following sub-section are the analogue to digital conversion, end point detection, template preparation and finally pitch detection. All of these were implemented in the Matlab environment.

2.1 Analog to Digital Conversion

This is performed by sampling the analog speech signal using an analog-to-digital converter. Since speech is typically bandlimited to about 4,000Hz, the speech signal needs to be sampled fast enough such that aliasing is avoided [6]. The
speech samples from the speakers were first recorded and converted into digitized speech using digital recorder. The recording requirements is shown in Table 1. The digital re- corder used as the analog to digital converter is the Goldwave software, which is shown in Fig. 1.

Table 1: The Recording Requirements

Recording Requirement

Description

Number of channels

Mono(1)

Sampling rate

12000Hz

Duration

One(1) second

Isolated words

Count Stop Down

Environment

Relatively low noise environment

Fig. 1: The Gold wave’s interface

2.2 End Point Detection

The end point detection method used is the linear pattern clas- sifier approach presented in [4]. End point detection and si- lence removal was carried out on the digitized speech. The algorithm uses statistical properties of background noise to make a sample as voiced or silence/unvoiced. The algorithm also uses physiological aspects of speech production for smoothening and reduction of probabilistic errors in statistical marking of the voiced or silence/unvoiced.
The two basic parameters used for determining the probability density function used for the statistical mapping are the mean

IJSER © 2015 http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 4, April-2015 3

ISSN 2229-5518

(μ) and variance (σ). The flowchart for the linear pattern clas- sifier is shown in Fig. 2, where, x and K are random variable and threshold respectively.

2.3 Template Preparation

The templates used for the development of an independent isolated word recognition system for this study were prepared using the average template method. For the independent iso- lated word recognition system developed for this study, each word was repeated 10 times by different speakers. After the end point detection, the features of each speech sample were extracted using Mel Frequency Ceptrum Coefficients (MFCC).
39 MFCC double delta coefficients were used per frame for each speech sample. Interested reader(s) can check the de- tailed procedures involved in our earliest article [10]. The block diagram for the average template method used in [5] and adopted from this study is shown in Fig. 3. Linear interpo- lation was used as the time normalization technique. This was achieved by using the expression; (1)
where, interp1 is a one dimension interpolation, Vq is the in-
terpolated speech sample X is the query points, X is the length of X , V is the speech sample to be interpolated and Method is the linear. The features of each speech sample after the end point detection were saved into the database.

2.4 Pitch Detection

The pitch detection was used to justify that the fundamental frequencies of each speaker varies and that it has a way of af- fecting the speech recognition accuracy. The Modified autocor- relation function (MACF) given in [11] was used as the pitch detection algorithm because it was more convenient for com- mon usage compared to others. The block diagram for MACF employed is shown in Fig. 4. The autocorrelation function is expressed in [11] as;
(2) Where, is the speech samples, N is the length of ana- lyzed frame, T is the number of autocorrelation points to be computed. The variable is called lag, or delay. Center-clipping was used to flatten the spectrum of the signal passed to the candidate generator. This is also expressed in [11] as;
START
Read the first 200ms of the signal sample
Read the sample from the first to last, calculate
Voiced sam- ple
Mark voiced sample as 1 and un- voiced sample as 0 to get an array of
0 and 1
Divide the array of 1 and 0 by
10ms overlapping windows
Label sample in each window either by 1 or 0 depending on the majority
Retrieve voiced samples by the only windows consist of 1
END
Unvoiced
Sample

(3) Fig. 2: The flowchart for the Linear Pattern Classifier [4]

IJSER © 2015 http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 4, April-2015 4

ISSN 2229-5518

3 DECISION LOGIC

After developing the speech recognition for this study, it was tested. The speech recognition was tested using the dynamic time warping to determine the optimum or the time- normalized distance. The input speech test was compared to the acoustic vectors of the three templates (count, stop and down). Templates already formed in the database for ten dif- ferent speakers were compared with the input speech test for the independent isolated word. Nearest Neighbour Classifica- tion (NNC) was used as the decision logic.
The Nearest Neighbour Classification is given as;
(4) where, is the recognized word, is the dynamic time warping (DTW), IST is the input speech test. The Nearest Neighbour Classification is explained by the block diagram given in Fig. 5.

Fig. 3: Block Diagram of the Average Template Method [5]

Fig. 5: The block diagram for the Nearest Neighbour Classifi-

cation

4 RESULT AND DISCUSSION

The results of the pre-processing stages are presented and discussed in this section. The digitized speech sam- ples of count, stop and down are given in Fig. 6.

Fig. 4: The block diagram for Modified Autocorrelation Func- tion [11]

The clipping threshold ( ) was set to be the maximum of the absolute value signal value. In order to enhance the pitch de- tection accuracy, median filtering was employed. The median filtering was carried out by using a Matlab function given as;
(4) where, medfilt1 is one dimensional median filter, X is the speech sequence, N is the order of the output, Y is the filtered
speech sequence.

Fig. 6: Speech Waveforms: (a) Count, (b) Stop and (c) Down

IJSER © 2015 http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 4, April-2015 5

ISSN 2229-5518

Fig. 7: Speech Waveforms after for End Point Detection for (a) Count, (b) Stop, (c) Down

Table 2 shows the threshold values (K) at which the voiced segment of the speaker utterance were extracted. It also shows the accuracy of the voiced segment obtained after end point detection. The Table 2 also shows whether the extracted voice segments of the utterances are correct or not when played back. The desired threshold value differs from one speech sample to the other even when a sample was repeated more than once. This is due to the fact that a speaker cannot utter the same word, the same way at different times. The results obtained show that the higher the threshold, the lower the accuracy of the voiced segment of the speaker ’s utterance. In order to pre- vent distortion of the voice segment, the threshold value was set at 0.3 for all the speech samples. The 0.3 threshold was used because at 0.3, the extracted voice segment of the utterances were all correctly obtained. Each end pointed utterance was obtained at a lesser number of samples. Figure 7 shows the speech waveform of count, stop and down after silence has been removed. The beginning and the ending of the speaker ’s utterance were detected when set at a threshold value of 0.3. Comparison between Fig. 6 and Fig. 7 shows that the silence in each speech sample was removed and the beginning and the ending of the speaker ’s utterance were detected. The speech samples were played back so as to confirm whether voiced segment were not eliminated with the unvoiced and silence portion of the speakers’ utterances.
After the template for the study had been prepared, it robust- ness and accuracy was tested by comparing the reference tem- plate or prepared template with one of the ten speakers for the speaker independent isolated word recognition system to be developed. The result obtained is presented in Fig. 8. The ob- tained results show that the count reference template and count
1 sample follows the same pattern with reduced local distance. The effect of preparing a master template is evident in the minimization of the local distance compared to when a single sample is used as the reference template.

Table 2: Voiced Segment after EPD at different Thresholds


Fig. 8: Comparison between the Count Reference Template and Count 1 Sample

The result of the pitch detection is presented in Table 3, which shows that the fundamental frequencies of ten speakers for
‘‘count’’, ‘‘stop’’ and ‘‘down’’ using MACF. The results show significant variation in the fundamental frequencies of ten speakers. However, the variation is not an indication of error in the template preparation but due to the fact that individual speaker has different pitch values. The result obtained, though varied for the 10 speakers fall within 91-209Hz with average fundamental frequency of 120.1Hz. The higher values are due to the age difference of the speakers and also the ambiguity of

IJSER © 2015 http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 4, April-2015 6

ISSN 2229-5518

peak detection in the MACF used for estimation of the pitch.
The recognition of the isolated words were achieved using the Nearest Neighbour Classification (NNC) as the decision logic. The decision logic used the DTW as the measure of global dis- similarity. The developed speech recognition system was ex- perimentally evaluated by collecting twenty (20) samples each for each word and were tested against the master templates already in the database using the 39 double delta acoustic vec- tors.
The overall average values of 23.3% and 76.7% were obtained as recognition accuracies before and after pre-processing stag- es respectively. The Fig. 9 and Fig. 10 shows the number of recognized speech samples for each word before and after the pre-processing stages. The comparative results obtained when pre-processing activities were employed and when they were not employed are presented in Fig. 9. The results show that the system recognition accuracy is better when pre-preprocessing stages were employed. This show that adequate pre- processing stage has significant effect on recognition accuracy of the word recognition system.

The variation observed in the recognition accuracies of the three speech samples buttresses the results presented in [13] that the ascent and the pitch differences have effects on the speakers’ utterances.

Table 3: Fundamental Frequencies of Speech Samples


Using MACF

Speaker Fundamental Frequency/Pitch (Hz)

Count Stop Down

1 99 91 134

2 178 136 124

3 169 209 121

4 162 209 125

5 181 127 119

6 112 122 93

7 92 92 93

8 92 92 91

9 92 92 94

20 With Pre-processing

10 95 91 92

Without Pre-processing

15

10

5

0

Count Stop Down

Speech Samples

Fig. 9: Comparative Recognition Accuracy with and without pre-processing activities

s

5 CONCLUSION

In the study, the effect of the pre-processing algorithms on accuracy of word recognition have been examined. The de- tailed information on the development of an isolated word recognition algorithms for the study was presented. The recognition was done using 39 MFCC delta coefficients and NNC as decision logic. The waveforms of speech samples after silence has been removed and the beginning and the ending of the speaker ’s utterance were detected at a threshold value of
0.3. The average recognition accuracy for sixty test input speech samples is 23.3% before pre-processing stages and
76.7% after the pre-processing stages. The effect of the prepar-
ing a master template is evident in the minimization of the local distance compared to when a single sample is used as the reference template. This aids easy warping of speech features and hence; improved recognition accuracy. The result of the study shows that both ascent and pitch has effect on word recognition accuracy.

IJSER © 2015 http://www.ijser.org

International Journal of Scientific & Engineering Research, Volume 6, Issue 4, April-2015 7

ISSN 2229-5518

REFERENCES

————————————————

Animasahun I.O. has B.ENG and M.ENG in Electrical and Electronics engi- neecing, communication option in Federal university of Technology Akure, Ondo State. E-mail: ianimasahun8619@gmail.com

Popoola J.J has B.S.C and M.ENG in the Federal University of Technology Akure, Ondo State and Ph.D in the University of Witwatersrand, South Afri- ca. E-mail: jidejulius2001@gmail.com

IJSER © 2015 http://www.ijser.org