A Knowledge based Approach Using Fuzzy Inference Rules for Vowel Recognition

Transcription

1 Journal of Convergence Information Technology Vol. 3 No 1, March 2008 A Knowledge based Approach Using Fuzzy Inference Rules for Vowel Recognition Hrudaya Ku. Tripathy* 1, B.K.Tripathy* 2 and Pradip K Das* 3 *1 Institute of Advanced Computer and Research, Prajukti Bihar, Rayagada (Orissa), India *2 School of Computing Sciences, VIT University Vellore , Tamil Nadu, India *3 Department of Computer Science & Engineering, Indian Institute of Technology Guwahati, North Guwahati (Assam), India hrudayakumar@hotmail.com, tripathybk@rediffmail.com, pkdas@iitg.ernet.in Abstract Automatic speech recognition by machine is one of the most efficient methods for man-machine communications. Because speech waveform is nonlinear and variant. Speech recognition requires a lot of intelligence and fault tolerance in the pattern recognition algorithms. Accurate vowel recognition forms the backbone of most successful speech recognition systems. A collection of techniques exists to extract the relevant features from the steady-state regions of the vowels both in time as well as in frequency domains. This paper is, introducing fuzzy techniques allow the classification of imprecise vowel data. By incorporating the acoustic attribute, the system acquires the capacity to correctly classify imprecise speech data input. Experimental results show that the fuzzy system s performance is vastly improved over a standard Mel frequency cepstral coefficient (MFCC) features analysis of vowel recognition. The speech recognition is a particularly difficult classification problem, due to differences in voice frequency (amongst speakers) and variations in pronunciation. Keywords Fuzzy Logic, Fuzzy Inference, Vowel, Speech Recognition 1. Introduction Automatic speech recognition by machine has been a part of science fiction for many years. The early attempts were made in the 1950s by various researchers. In 1952, Davis Biddulph and Balashek [1] designed the first isolated digit recognizer for a single speaker at the Bell Laboratories. This system used a simple pattern matching method with templates for each of the digits. Matching was performed with two parameters: a frequency cut based on separating the spectrum of the spoken digit into two bands and a fundamental frequency estimated by zero-crossing counting. The 1970s and 1980s were very active periods for speech recognition with a series of important milestones: Pattern recognition algorithms were applied for the template-based isolated word recognition methods. Continuous speech from large vocabularies was understood based on the use of high-level knowledge to compensate for the errors in phonetic approaches. Speech analysis method based on Linear Predictive Coding (LPC) was used instead of conventional methods such as FFT and filter banks. Statistical modelings such as the HMMs (Hidden Markov Model) were developed for continuous speech recognition. The neural networks (back propagation. learning vector quantization) with efficient learning algorithms were proposed for speech pattern matching. Vowels are generally well defined in the spectral domain. As such, they contribute significantly to our ability to recognize speech, both by human beings and speech recognizers. For vowels, the speech behavior can be considered as a point that moves in parameter 51

2 A Knowledge based Approach Using Fuzzy Inference Rules for Vowel Recognition Hrudaya Ku. Tripathy, B.K.Tripathy and Pradip K Das space as the articulatory system changes. Standard HMMs using cepstral coefficients with their derivatives cannot effectively model the trajectories especially for vowels [2]. In recent years the speech recognition technology have begun to enter the real world in our life. More and more advanced algorithms were adopted in this area. The paper [3] presented an improved vowel detection and segmentation scheme. The vowel region of the wave file is removed after perceptually checking the vowel in question. This wave file is converted into data text file and is filtered for DC shift to remove internal machine noise. LP analysis followed by computation of cepstral coefficients and weighing is done to form the feature vectors. A series of experiments were reported by choosing vowel segments carefully based on perception. It was reported that the vowel recognition scores were better than the standard procedures. In this paper the Fuzzy logic techniques have been applied to classification of vowel for speech recognition and this field is growing and developing very fast. 2. Introduction to Speech Sounds 2.1 Speech Production Speech sound is produced by a set of wellcontrolled movements of various speech apparatus. Figure 1: shows a schematic cross-section through the vocal tract of the apparatus. The vocal tract is a primary acoustic tube, which is the region of the mouth cavity bounded by the vocal cords and the lips. As air is expelled from the lungs, the vocal cords are tensed and then caused to vibrate by the airflow. The frequency of oscillation is called the fundamental frequency, and it depends on the length tension and mass of the vocal cords. During this process, the shape of the vocal tube is changed by different positions of the velum, tongue, jaw and lips [4]. The average length of the vocal tract for an adult male is about 17cm. and its cross-section area can vary in its outer section from 0 to about 20cm 2. Therefore, the vocal tract, as an acoustic resonator, will determine variable resonant frequencies by adjusting the shape and size of the vocal tract. The resonant frequency is called the formant frequency or simply formant. The nasal tract is an auxiliary acoustic tube that can be acoustically cooperated with vocal tract to produce nasal sounds. Not only adjusting the shape of the vocal tract, but also the type of excitation produces various speech sounds. Besides the airflow from the lung, the excitation could come from some other sources: the fricative excitation and whispered excitation [5]. Figure 1. Schematic view of the human speech apparatus 2.2 Fundamental Speech Recognition Techniques Classification of Speech Recognizer Automatic speech recognition cm be classified into a number of different categories depending on different issues: 1. The manner in which a user speaks. Usually there are three recognition modes based on the speaking manner: Isolated word recognition: The user speaks individual words or phrases from a specified vocabulary. Isolated word recognition is suitable for command recognition. Connected word recognition: The user speaks fluent sequence of words with small spaces between words, in which each word is from a specified vocabulary (e.g.. zip codes. phone numbers). Continuous speech recognition: The speaker can speak fluently with a large vocabulary. 2. The number of users: Speaker dependent: The users of a recognition system only consist of a single speaker or a set of known speakers. Speaker independent: arbitrary users will use the ASR system in this case. 52

3 Journal of Convergence Information Technology Vol. 3 No 1, March 2008 Speaker adaptive: The system will customize its response to each individual speaker while it is in use by the speaker. 3. The size of the recognition vocabulary: A small vocabulary system only provides recognition capability for a small amount of words A large vocabu1ary system is capable of recognizing words among a vocabulary containing up to 1000 words. 3. Fuzzy Logic 3.1 Background Fuzzy sets were introduced by Zadeh [6] in 1965 as a new way to represent and manipulate data with uncertainty and fuzziness. In the old paradigm, fuzziness was considered unfavorable because of the expectation for scientific precision and accuracy. However, a fuzzy interpretation of data is a natural and intuitively possible way to formulate and solve a lot of problems in our everyday life. For example, expressions with uncertainty like "hot coffee", "heavy objects", and "warm weather" are fuzzy interpretations. Although both fuzzy sets and statistical theory can deal with uncertainty, fuzzy sets are quite different from statistical models in some ways. Probabilities represent the likelihood of a certain event with a distribution among all the events, while a fuzzy set represents the applicability of the element to the set. In another word, the fuzziness provides more uncertainty that can be found in the meanings of many words from human's thinking. 3.2 Fuzzy set and Fuzzy Logic Fuzzy sets are a super-set of classical sets. In a fuzzy set, each element is associated with a real value, which represents the degree of membership of the element in the closed unit interval [0,1]. However, in classical crisp sets, al1 element can only be classified as "0" or "1". When al1 elements in a set have either complete membership or complete non-membership, the fuzzy set reduces to a crisp set [7]. Suppose a fuzzy set A is a subset in space X that admits partial membership. It is defined as the ordered pair A = {x, m A (x)}, where x X and 0 m A (x) 1. Every fuzzy set consists of the three parts: a horizontal axis x specifying the population of sets, a vertical membership axis m A (x) which specifies the membership degree of each element and the surface itself to provide a one to one connection between the elements and their corresponding membership degree Fuzzy System Fuzzy systems use fuzzy set theory to deal with fuzzy or non-fuzzy information. Generally, a fuzzy system consists of a fuzzification subsystem. a fuzzy inference engine. a fuzzy rule base and a defuzzifier as shown in Figure 2. The fuzzy rule base and fuzzy inference engine is the core of the fuzzy-rule-based system. A fuzzy rule can be expressed by a set of fuzzy inference rules in the form of "IF x is A THEN y is B" [8], [9]. The inference engine then implements a fuzzy inference algorithm to determine the fuzzy output from the inference rules and the inputs. Finalized fuzzification follows the central stage of fuzzy system functioning fuzzy inferencing. This phase (Figure 2.) understands usage of knowledge base i.e. execution of the aggregation of fuzzy production system fuzzy premises rule application, adequate to context of fuzzy inferencing system model [10]. Note that a given input may simultaneously be a member of more than one set within a single fuzzy region. The inference engine interacts with the rule base and uses the inputs to determine which rules are applicable. The outputs are a set of fuzzy sets defined on the universe of possible outputs, which will be defuzzified to generate crisp outputs. Fuzzification Subsystem Inference Engine Fuzzy Rule Base Defuzzification System Figure 2. A typical fuzzy rule based system Frequent practice of this way of inferencing, assuming fuzzy logic systems for decision support and different systems control, one can find in literature. This was the original idea, exploited in this paper. Such system could be exploited for speech recognition because of non-numerical parameters. Conceptual extension of classic inferencing and control systems consists of absence of analytic description of such systems. First approaches to an extension of control systems based on Zadeh's concept of fuzzy sets origin from Mamdani [8], who has introduced a fuzzy logic controller which contained control algorithm based on simple rules. Approximative reasoning of such fuzzy system converts knowledge represented by incomplete (fuzzy) information and fuzzy rules into a non-fuzzy (numeric) outputs. In order to model human reasoning 53

4 A Knowledge based Approach Using Fuzzy Inference Rules for Vowel Recognition Hrudaya Ku. Tripathy, B.K.Tripathy and Pradip K Das mechanisms, Lofti Zadeh has introduced the fuzzy extension of conventional inferencing systems (fuzzy logic systems - FLS) that, besides quantitative aspects, have included the logic of inexact, incomplete information, operation and inferencing rules, also. In order to combine with heuristic formulation inside such systems, so called, if-then rules, numerical values of mathematical descriptions had to be symbolically interpreted. appropriately found on the basis of the succeeding pitch start position. Again, it will follow the same method to select the next pitch period. In this way, it will store 8 to 10 pitch period s raw data in a text file according to the steady part of the waveform. 4. Methodology and Preprocessing The methodology of proposed knowledge based fuzzy inference rule is shown in Figure 3. The speech waveform data consists of 3700 pitches of 200 utterances containing 10 repetitions of 5 different vowels spoken by 20 male speakers. All the speakers are from different parts of India. The vowels are recorded with a carrier sentence I say aaaaa now, I say oooo now, etc. as per IPA standard. All the speech utterances used in the present study are recorded in an air-conditioned lab with presence of number of students. Recording software was used for recording with a sampling rate of 22050Hz, mono channel and 16-bit resolution. Input Vowel Convert to Raw data Fuzzy Inference Algorithm Recognized Vowel Pitch extraction Acoustic Analysis to form Fuzzy set Figure 3. Block diagram of proposed Knowledge based Approach Using Fuzzy Inference Rules for Vowel Recognition The raw speech data is extracted from the respective recorded wave files and stored in a text format. Further the raw data of the wave file was normalized with an absolute maximum value. An algorithm was developed to accurately detect a pitch period on the basis of the highest positive sample value in the data and scanning to find the lowest positive value preceding the located highest value. The position of the selected lowest positive sample will be the pitch start position. Similarly, after getting pitch start position, data is scanned forward from first highest positive sample value to reach to the next highest positive sample value. The pitch end position is Figure 4. Formulation of Fuzzy set by analyzing of Pitch period of a speech. Calculate the number of available positive going curves in an each pitch period and find out the number of peaks present in all such curves. The data corresponding to the pitch period is isolated. Next, compute the number of peaks and the number of samples of first positive going curve of the pitch period as shown in Figure 4. This is done for all vowels. The algorithm so designed to ensure that pitch periods in which there are less than 4 samples in the first positive going curve is ignored for computation. This takes care that most invalid pitch periods are removed from consideration. 4.1 Acoustic analysis using Fuzzy inference We describe the vowel recognition classifier algorithm based on the number of samples and the number of peaks present in the first positive going curve. Along with this the information on the total number of peaks located within on pitch period is also used. The proposed Fuzzy inference algorithm classifies the spoken English vowels based on the hierarchy scheme. As initially, the algorithm classifies a given input vowel into two groups namely {/o/,/u/} and {/a/,/i/,/e/}. From the first group /o/ and /u/ are recognized separately. Next, it will proceed to differentiate the given vowel between /a/ and {/i/,/e/}. Finally, the vowel is either classified as /i/ or /e/. This tree type heuristic classification is based on the acoustic analysis of the vowel waveforms. 54

5 Journal of Convergence Information Technology Vol. 3 No 1, March 2008 Heuristically it was verified by plotting the graph of all utterances of large number of pitch periods, which falls in a particular range for respective vowels. For example, as in vowel /a/ the percentage of more number of samples in first positive going curve was in between 1 to 19, percentage of peaks in the same curve was from 1 to 3 and the total number of peaks in all positive going curve of a whole pitch period was also from 1 to 8 as shown in Figure 5. Number of samples Number of peaks Number of samples in first positive curve of a vowel '/a/' Category of Samples Total number of peaks in all positive curve of a pitch of a vowel '/a/' Category of peaks 9 From Figure 8, the recognition score for /u/ and /o/ is found to be 99% and 93% respectively. Vowel /a/ has been correctly recognized with an accuracy of 92%. It is observed that in case of /e/ and /i/ the recognition score is low 82% and 86% respectively. if ((number of samples in first positive curve >=19) && (number of peaks in first positive curve <=3) && (total no of peaks in all positive curve <=8)) { if (number of samples in first positive curve <=26) printf ( Vowel is O ); else printf( Vowel is U ); } if ((number of samples in first positive curve <19) && (number of peaks in first positive curve <=3) && (total no of peaks in all positive curve <=10)) printf ( Vowel is A ); if ((number of peaks in first positive curve <=4) && (total no of peaks in all positive curve >10) && (total no of peaks in all positive curve <=20)) printf ( Vowel is E ); if ((number of samples in first positive curve >=10) && (number of peaks in first positive curve >4)) printf ( Vowel is I ); Figure 6. Fuzzy inference algorithm for vowel classification. Figure 5. Graphs of Fuzzy set analysis of vowel /a/. Similarly, computation of characterization is done for the remaining vowels /i/,/e/,/o/ and /u/. On the basis of information extracted from 3700 pitch periods of 5 vowels uttered by 4 ethnically different 20 Indian speakers. 4.2 Results On the basis of the logic developed in Fuzzy inference algorithm as explained in Figure 6, a set of 3700 pitch periods of 5 vowels spoken by 20 adult male speakers were used to test the classification system for its correctness. Figure 7, shows the output for the recognition of the 5 vowels. It should be noted that some pitch period were removed from the classification scheme as the number of samples detected in the first positive going curve is found to be less than 4 as described earlier. Experimental results show in comparison that the fuzzy system s performance is vastly improved over a standard Mel frequency cepstral coefficient (MFCC) features analysis of vowel recognition as in [3]. Figure 7. Output for the recognition of the vowel 55

6 A Knowledge based Approach Using Fuzzy Inference Rules for Vowel Recognition Hrudaya Ku. Tripathy, B.K.Tripathy and Pradip K Das [7] Hui Ping, Isolated Word Speech recognition Using Fuzzy Neural Techniques, A Thesis Submitted to the College of Graduate Studies and Research, University of Windsor Windsor. Ontario. Canada, [8] Lynn Yaling Cai, Hon Keung Kwan, "Fuzzy classifications using fùzzy inference networks", IEEE Transactions on Systems, Man and Cybernetic-Part B: Cybernetics, Vol. 28, No. 3, June 1998, pp [9] Hon Keung Ktvan. Yaling Cai, and Bin Zhang, "Membership function learning in fuzzy classification". International Journal of Electronics, Vol. 74, No. 6, pp Conclusion Figure 8. Vowel recognition accuracy Fuzzy systems based knowledge in speech recognition that finds useful practical applications in situations have been considered in this paper. An identification of basic acoustic parameters, established Fuzzy inference rule and the way of its classification of vowels are also illustrated. 6. References [1] K. Davis. R. Biddulph. and S. Balashek. "Automatic recognition of spoken digits". Journal Acoustic Society America, 1952, 23: pp [2] Y. Gong, Stochastic Trajectory Modeling and Sentence Searching for Continuous Speech Recognition, IEEE Trans. SAP, Vol. 5, No. 1,January (1997) [3] K. Saravanakumar, Hrudaya Ku. Tripathy, Pradip K. Das, An Improved Wave Segmentation Scheme for Vowel Recognition, Proceedings of the National Conference on Communication Technologies (NCCT-2006), Mepco Schlenk Engg. College, Tamil Nadu, (2006), [4] Lawrence Rabinar. Bing-hwang Juang, Fundamentals of speech recognition. Prentice Hall. Englewood Cliffs [10] Miloš Manić, Dragan Cvetković, Momir Praščević, Intelligibility Speech Estimation Using Fuzzy Logic Inferencing, The scientific journal FACTA UNIVERSITATIS, Vol. 1, No 4, 1999, pp Authors bio. Hrudaya Ku. Tripathy is presently working as Asst. Professor in the Dept. of Computer Science & Engineering at Institute of Advanced Computer & Research, Rayagada, Orissa, India. He has 10 years of teaching experience in UG/PG courses. He has M.Tech from Indian Institute of Technology Guwahati, and pursuing his Doctorate under Berhampur University, Berhampur Orissa, India. B.K.Tripathy is presently working as Professor in the School of Computing Science at VIT University, Vellore, India. Having around 27 years of teaching experience in UG/PG courses. He has M.Tech from University of Poona, Pune, India. and completed his PhD. He has published more than 50 research papers in national and international journals. Pradip K. Das is presently working as Asst. Professor in the Dept. of Computer Science & Engineering at Indian Institute of Technology Guwahati. Having around 20 years of teaching experience in both UG/PG classes. He has M.Sc from Delhi University and completed his PhD. He has published more than 30 research papers in national and international journals. [5] Joseph P. Campbell, JR., "Speaker recognition: A tutorial". Proceedings of IEEE, Vol. 85. No. 9. September [6] L. A. Zadeh. Fuzzy sets, Information Control, pp