Myoelectric Signal Classification for Phoneme ... - Semantic Scholar
694
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 4, APRIL 2007
Myoelectric Signal Classification for Phoneme-Based Speech Recognition Erik J. Scheme, Bernard Hudgins*, Senior Member, IEEE, and Phillip A. Parker, Senior Member, IEEE
Abstract—Traditional acoustic speech recognition accuracies have been shown to deteriorate in highly noisy environments. A secondary information source is exploited using surface myoelectric signals (MES) collected from facial articulatory muscles during speech. Words are classified at the phoneme level using a hidden Markov model (HMM) classifier. Acoustic and MES data was collected while the words “zero” through “nine” were spoken. An acoustic expert classified the 18 formative phonemes in low noise levels [signal-to-noise ratio (SNR) of 17.5 dB] with an accuracy of 99%, but deteriorated to approximately 38% under simulations with SNR approaching 0 dB. A fused acoustic-myoelectric multiexpert system, without knowledge of SNR, improved on acoustic classification results at all noise levels. A multiexpert system, incorporating SNR information, obtained accuracies of 99% at low noise levels while maintaining accuracies above 94% during low SNR (0 dB) simulations. Results improve on previous full word MES speech recognition accuracies by almost 10%.
MES (which is generated by muscles during contraction) is immune to acoustic noise, data collected from facial muscles during speech could be used to augment conventional acoustic recognition. This is particularly convenient for fighter pilots as electrodes can be placed inside the oxygen mask. MES classification is also a good complement to acoustic speech recognition (ASR) because sounds that are acoustically similar are often formed using reasonably different musculatures. Herein, Chan’s approach is expanded to classify words at the phoneme level. This provides a more detailed analysis by identifying the phonemic makeup of each word. Phoneme based recognition also allows for more rapid expansion of the recognizable vocabulary.
Index Terms—EMG, HMM, multi-expert systems, myoelectric signal, speech recognition.
II. BACKGROUND
I. INTRODUCTION
S
PEECH recognition has become a widely researched field covering tasks such as autonomous transcription on a home computer to advanced military and security applications. One proposed function is the control of secondary tasks in military fighter jets. Due to the complexity of their instrument panels and controls, pilots are forced to place a large portion of their focus downwards. It is desirable to implement a speech recognition system that would alleviate some of this dependence and enable more concentration to be shifted towards “heads-up” flying. However in order for speech recognition to be a usable tool in such a critical function accuracy rates must be exceptional. Due to the noisy environment in such aircraft, conventional speech recognition is limited. Although much research has been done to improve acoustic results under noisy conditions [1]–[3], it is beneficial to consider alternate approaches. Chan [4] proposed a multiexpert system using MES classification combined with acoustic recognition. Because the
Manuscript received November 25, 2005; revised September 16, 2006. This work was supported in part by the NSERC under Discovery Grant 171368-03 and Discovery Grant 4445-04 and in part by the New Brunswick Innovation Fund RAI program. Asterisk indicates corresponding author. E. J. Scheme is with the Institute of Biomedical Engineering and the Department of Electrical and Computer Engineering, University of New Brunswick, Fredericton, NB E3B 5A3, Canada. *B. Hudgins is with the Institute of Biomedical Engineering, University of New Brunswick, PO Box 4400, Fredericton, NB E3B 5A3, Canada (e-mail: [email protected]). P. A. Parker is with the Institute of Biomedical Engineering, University of New Brunswick, Fredericton, NB E3B 5A3, Canada. Digital Object Identifier 10.1109/TBME.2006.889175
The first appearances of MES based voice or speech prostheses occurred in the mid 1980s. Morse and O’Brien [5] used four channels of MES data (three recorded from the neck and one from the forehead) to obtain a 97% classification rate for a 2-word vocabulary. Later, Morse et al. reported accuracies of 57% [6] and 60% [7] for 10-word vocabularies. Sugie et al. [8] suggested the use of a voice prosthesis for discriminating between the five Japanese vowels. Using an automaton, they were able to achieve an average classification rate of approximately 60% using three channels of MES recorded from sites around the mouth. Chan et al. [4] proposed a method of speech recognition using a multiexpert system combining both acoustic and MES recognition. In a preliminary study using 5 channels of facial MES data, Chan was able to obtain an accuracy of approximately 85% and 87% when classifying the English words “zero” through “nine” for two subjects. When combined with an acoustic expert, this system proved to greatly improve recognition rates in noisy environments when compared to a purely acoustic classifier. In 2003, Manabe et al. [9] used MES data recorded from three facial muscles during unvoiced speech to perform what they called “mimed speech recognition.” Three subjects mouthed the five Japanese vowels for periods of 10 s, and 400 ms portions were later extracted to form training and test sets. Using a three-layer neural network, Manabe obtained an average classification rate of over 94%. The vowels were not taken contextually from speech and were mouthed for extended periods, effectively eliminating any transient or coarticulatory effects which may have occurred in conversational speech. Although research in this field has been relatively limited, results have shown that significant information exists in the MES
0018-9294/$25.00 © 2007 IEEE
SCHEME et al.: MES CLASSIFICATION FOR PHONEME-BASED SPEECH RECOGNITION
695
IV. SEGMENTATION
Fig. 1. Electrode placement.
of articulatory muscles. It has been shown that speech information is present at the word level and from separate vocalizations of vowels. Herein, we confirm the presence of phonemic data in the MES recorded from spoken words. III. METHOD As in Chan’s study, five channels of MES data were recorded using Ag-AgCl Duotrode bipolar electrode pairs placed over five articulatory muscles of the face: the levator anguli oris, the zygomaticus, the platysma, the depressor anguli oris, and the anterior belly of the digastricus. Unlike in Chan’s work, electrodes were not embedded in a pilot’s oxygen mask because the mask severely degraded the quality of acoustic speech needed for phonemic segmentation. Instead, electrode pairs were placed directly on the speakers face (Fig. 1), and taped as required to minimize any extraneous movement during speech. The subject’s skin was cleaned using alcohol swabs to remove any dirt or makeup, and electrode gel was added to the electrodes to minimize impedance at the skin-electrode interface. An Ag-AgCl Red-Dot electrode (3M, 2259) was also placed on the back of the neck to provide a common reference. All signals were then sampled at 5 kHz using a simultaneous sample and hold device (ComputerBoards, CIO SSH16) and a 16-bit analog to digital board (Measurement Computing, PCIMDAS1602/16). Four English speaking subjects (3 male, 1 female) each performed 9 sets of 5 repetitions of the words “zero” through “nine.” Subjects were encouraged to rest between sets and the words were presented in random order to minimize anticipatory articulation. It was found, in early trials, that when a subject knew the next word to be spoken, there was a tendency to preposition the facial muscles earlier than in natural speech. All training and testing data were collected during one session due to subject availability. A multisession experiment would require electrode placement to be carefully reproduced. However, the classifier exploits interchannel patterns and it is, therefore, believed that reasonable amounts of intersession variability are tolerable and trained models can be used across sessions. A mask embedded with electrodes, such as that used by Chan [4] but causing less acoustic distortion, could sufficiently minimize electrode placement variability.
In order to train a phoneme classifier, it is necessary to first have speech that is segmented into phonemes. However, there is currently no published information about the phonemic segmentation of MES. Also, considering that the MES of two individuals is distinct [10] (making it impossible to form a general MES phoneme model across subjects) it quickly becomes apparent that directly segmenting the MES is a challenge. Morse and O’Brien [5] showed that a slight delay exists between the onset of MES from facial muscles and the appearance of its associated acoustic speech. This is due to the fact that the muscles in the vocal tract must be contracted prior to the generation of sound. A slight delay is also apparent due to the time required for sound to propagate to a microphone. It was, therefore, decided to segment the acoustic data, apply a negative time shift and relate the corresponding segments/labels to the associated MES. Current automatic phoneme segmentation algorithms [11] correctly segment approximately 74% of phonemes without over-segmenting. The incorrectly labeled data is not only skewed, but false phonemes can be inserted, and correct phonemes disregarded. Therefore, even when an automatic labeler is implemented, a manual verification of each and every result is required [12]. Both the training and test data for this experiment was, therefore, manually segmented. Segmentation was facilitated by using custom Matlab software that allowed the user to view spectrographic and formant information as well as use audio cues for labelling phonemes. During initial segmentation trials, phonemes were chosen to be continuous within a word, starting immediately after the preceding phoneme ended. It was later found that results could be improved when segments were labelled more stringently to avoid incorporating coarticulatory data. This effect exists during the natural flow of speech at the time when two phonemes are strung together, thereby changing their characteristics. It is possible that coarticulation effects may even be exaggerated in MES due to anticipatory muscle contractions which occur while prepositioning muscles for articulation. The acoustic speech data was, therefore, manually re-segmented with focus on selecting only those parts caused purely by the desired phoneme (Fig. 2). Again, time-shifted versions of the resulting labels were applied to the MES data. Sixty percent of each speaker’s data set was used for training, while the remaining 40% was used for testing. It was found that increasing the number of repetitions in training beyond 27 (60% of 45) produced diminished returns in classification accuracy. The test set for each speaker, therefore, consisted of 18 repetitions of each word. Once the training words were chosen, the associated phonemes were sorted into input training vectors for each HMM phoneme model. The phonemic representation of the words as used herein is shown in Table I. Rather than phonemes, as used herein, some systems classify triphones which incorporate information from the phonemes preceding and following a particular phone. This has been shown to increase performance of recognition systems, but at the cost of greatly increasing the size of the collection of combinations classifiable segments. There are
696
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 4, APRIL 2007
TABLE II CLASSIFICATION RESULTS FOR ACOUSTIC DATA
Fig. 2. Acoustic segmentation of the word “zero” with negative time shift applied to MES data.
TABLE I PHONEMIC COMPOSITION OF TEST DATA
of 3 phonemes. While the English language limits the number of allowable combinations, there are still thousands of triphones compared to the 44 phonemes. Many research projects use a selection of words chosen to minimize the number of spoken triphones. In the 18 phoneme vocabulary used in this research, there are 30 different triphones. V. CLASSIFICATION The HMM classifier used during this research was developed using the H2M toolbox for Matlab [13]. Five state, left-to-right HMM were trained with single Gaussian mixture observation densities using the Forward-Backward algorithm [14] for each phoneme. Because each person’s MES is unique, speaker dependent models were required. Features for the acoustic classifier consisted of the root-mean-square value of each frame and the first 12 Mel-cepstral frequency coefficients (MFCC). Because of their close approximation to the human perception of
sound, MFCC have become widely used in ASR. However, because the Mel-scale warping is tailored to the human perception of acoustic speech, it would initially appear that this warping of the frequency spectrum might not apply to MES. However, Chan [4] found that MFCC performed better than other feature sets when classifying facial MES recorded during speech. The same feature set was, therefore, used to represent each of the five channels of data for the MES classifier. While Chan [4] used frames of 64 ms, window lengths herein were restricted due to the small size of certain phonemic segments. It was necessary to constrain the frame length and increment in order to obtain a reasonable number of frames when representing short phonemes. Classification performance suffered when greater window lengths were chosen because information from adjoining phonemes became included in the frame. While these parameters were investigated separately for acoustic and MES data (with frame lengths from 5 ms to 100 ms), the configuration that provided the best classification accuracy for both consisted of frame lengths of 12.8 ms, overlapped by 1.6 ms. Phoneme classification decisions were made using the maximum likelihood output of the Viterbi algorithm [14]. For a given segment of speech (a word), the phoneme classification results were then compared to those in a dictionary and the segment was assigned a score based on the number of matching phonemes. The dictionary consisted of a list of the test words (“zero” through “nine”) along with their phonemic breakdowns. The dictionary word which corresponded to the highest score was then selected as the recognized word. It is possible for two words to return the same maximum score. For example, the phonemes “f”-“i”-“r” would return an equal score for the words “four” (“f”-“oh”-“r”) and “five” (“f”“i”-“v”). In this case, a likelihood score was determined by summing the likelihoods associated with each word. This was done by summing the original likelihoods for each phoneme, as obtained from the classification process. The maximum likelihood would then be selected and the associated word chosen. VI. RESULTS The acoustic data was classified using the discussed HMM classifier. The results of the classification are shown in Table II. While the average phoneme classification accuracy is slightly over 91%, the overall word accuracy is over 99%. This illustrates one of the benefits of using a segmental approach to speech classification. The combination of multiple phonemes
SCHEME et al.: MES CLASSIFICATION FOR PHONEME-BASED SPEECH RECOGNITION
697
TABLE III CLASSIFICATION RESULTS FOR MES DATA
Fig. 3. Effect of offset between acoustic and MES segmentation boundaries on overall classification accuracy.
into one word (following a structured grammatical format) allows a certain level of forgiveness. If a phoneme within a word is misclassified, there may be sufficient information provided by the other phonemes to correctly identify the word. Because there is a delay between the onset of MES activity and the appearance of acoustic speech [5], a time shifted variation of the acoustic segmentation was used to segment the MES data. The choice of offset is crucial because it is necessary to have the exact MES information that corresponds to the acoustic phoneme being spoken. Chan [4] applied a shift value of 500 ms to make sure that all MES data was included in the overall time window representing a spoken word. However, for phoneme based recognition it is necessary to further scrutinize this delay. During speech, natural silences exist between words, allowing some room for error in the location of a selection window. However, because phonemes within a word are typically strung together without pause, the window must (optimally) envelope the desired phoneme without encompassing any extra speech. To test this, classification accuracy was measured as a function of offset/delay. The average effect is shown in Fig. 3. It can be seen that a distinct peak exists around the 100 ms mark. This implies that the optimal correlation between the acoustic and MES data occurs at a delay of 100 ms. This is consistent with the literature for typical electro-mechanical delays in muscle [16]. Having selected the optimal offset, classification of the MES was performed. Table III shows the obtained results. These results show an almost 10% improvement over results obtained by Chan [4] in a study of the effect of temporal misalignment on MES-based speech recognition. Chan used full word HMM classification to attain an accuracy of approximately 85% using the same 10-word vocabulary.
In order to better demonstrate the merits of using MES speech recognition, it is beneficial to investigate its performance in acoustically noisy environments. Varying levels of white Gaussian noise were added to the collected acoustic data so that the SNR (calculated using a ratio of signal power) resulting from the most severely degraded word sample, on a per speaker basis, was approximately 0 dB. This was determined empirically so that acoustic phoneme recognition rates dropped to approximately a priori at the highest noise level. The other levels of noise were distributed between clean (the unaltered collected data was designated as the “clean” class) and the 0 dB SNR. HMM were trained using the unaltered data and classification was performed on all levels of the noise-corrupted data. No acoustic noise filtering or cancellation was implemented in order to better illustrate the effect of noise on acoustic classification rates. It can be seen in Fig. 4 that the acoustic classification accuracy falls to 8% at 0 dB, approximately the a priori rate of 6% for 18 phonemes. For the purpose of this investigation, the Lombard effect has been assumed to be negligible. The Lombard effect is the change in the way a person speaks when in acoustically noisy conditions [17]. It has been suggested that changes in power, shifting of formant frequencies, and changes in formant band width (among others) can occur, however Chan [4] found that the effect did not appreciably change facial MES at the noise levels used herein. Because, under this assumption, the MES is essentially immune to acoustic noise, the collected MES data was assumed to be constant across all levels of acoustic noise. This is depicted in Fig. 4 as a horizontal line for MES classification rates across varying acoustic noise levels. Note that the MES classification rate is inferior to the acoustic rate for only low levels of noise. When higher levels of noise are introduced, the MES becomes a better source of information for phoneme recognition. Because the classifiers perform differently in varying noise levels, it is desirable to exploit their relative strengths. A multiexpert system combines the opinions of two or more classifiers, yielding an overall classification decision which is a function of the individual verdicts of its formative experts. Chan [4] describes several ways of forming multiexpert systems, incorporating various levels of abstractness (majority vote, Bayesian belief, Dempster-Shafer, etc). In particular, Chan introduces a plausibility method of combining experts using mathematical theory of evidence. The method assigns weights to the formative classifiers based
698
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 4, APRIL 2007
TABLE IV AVERAGE PHONEME ACCURACY (%) VERSUS SNR
TABLE V AVERAGE WORD ACCURACY (%) VERSUS SNR
Fig. 4. Phoneme (top) and word (bottom) classification accuracies of the acoustic, MES, and multiexpert systems (with and without knowledge of acoustic SNR) versus acoustic SNR.
on their ability to discriminate information. The algorithm is particularly beneficial in applications where the SNR cannot be measured, because it is able to assign beliefs based on the distribution of likelihoods, without empirical knowledge of performance. A plausibility based multiexpert system was implemented, combining the likelihoods of the acoustic and MES classifiers. The phoneme corresponding to the maximum resulting likelihood was selected as the multiexpert result. Fig. 4 shows that the fused system [denoted as “Multi-Expert (No SNR)”] outperforms the acoustic classifier alone at all noise levels. This confirms that the MES classifier is able to discern information orthogonal to that of the acoustic classifier, and can be used to improve accuracy even in low noise environments. At high noise levels, however, the classification accuracy of the multiexpert system drops below that of the MES classifier. An optimal fusion system would produce accuracies equal to or greater than that of the most accurate classifier at all noise levels. When the acoustic SNR ratio is known, it is possible to use heuristic information about classifier accuracies in the presence of noise to determine optimal weightings for a multiexpert combination equation. The experts were, therefore, combined using a weighted sum of their outputs, where the weights were a function of each classifier’s accuracy (the ensemble mean across all subjects) in varying levels of acoustic SNR. Matlab’s curve fitting toolbox was used to approximate a continuous weighting curve for the
acoustic classifier (shown in (1)). Because the phoneme accuracy of the MES classifier is independent of acoustic noise, its weight was a constant,
where (1) , was calculated by The multiexpert likelihood, weighting the acoustic classifier likelihood, , and adding it to a weighted version of the MES classifier likelihood, [shown in (2)]. The phoneme with the highest likelihood was then selected in the same way it was for each individual classifier (2) While this system requires knowledge of the SNR, it can be seen in Fig. 4 that it matches or outperforms both the acoustic and MES classifier at all noise levels. These results are also shown below in Tables IV and V. VII. CONCLUSION Phoneme based acoustic and MES speech recognition systems were presented. MES were extracted from 5 articulatory muscles in the face as a secondary source of speech information. The digits from “zero” through “nine” were used as the training
SCHEME et al.: MES CLASSIFICATION FOR PHONEME-BASED SPEECH RECOGNITION
and test vocabulary. Manual segmentation was performed in order to demonstrate the accuracy of the phoneme recognition without performance degradation due to errors in automatic segmentation. The acoustic system outperformed the MES system at low noise levels (99% to 95%), however its performance significantly deteriorated at lower SNRs, while the acoustic noise immune MES system upheld its accuracy. It was shown that a multiexpert system combining the opinions of both classifiers, without knowledge of SNR, outperformed the acoustic classifier at all noise levels. A multiexpert system, incorporating information about SNR, outperformed both classifiers at all noise levels, obtaining word accuracies above 99% at low noise levels and above 94% at noise levels approaching 0 dB. In addition, MES classification at the phoneme level yielded accuracies almost 10% better then those obtained by Chan [4] using full word recognition for the same vocabulary. MES phoneme recognition improves upon previous wordbased MES systems by enabling the addition of new words without further training. If an additional word were to be added to a word-based system, a new model would first need to be trained and tested. Using the method discussed herein, any word comprised of the current phonemes (18 have been trained out of the approximately 44 in the English language [15]) can be added to the dictionary without additional training. A system containing all 44 phonemes would be able to be expanded to classify any word in the English language by simply adding that word to the dictionary. A maximum of 44 HMM would, therefore, be necessary, whereas the number of HMM would increase at a 1:1 ratio with the number of words added to a full word recognition system. Future research will investigate the expansion of the phoneme database, as well as the feasibility and accuracy of real-time acoustic-MES speech recognition using automatic segmentation of speech. Alternate sources of noise, such as electromagnetic and effects of gravity on speech will also be investigated.
REFERENCES [1] B. Widrow, “Adaptive noise cancelling: principles and applications,” Proc. IEEE, vol. 63, no. 12, pp. 1692–1716, Dec. 1975. [2] M. R. Sambur, “Adaptive noise canceling for speech signals,” IEEE Trans. Acoust., Speech, Signal Process., vol. ASSP-26, no. 5, pp. 419–423, Oct. 1978. [3] B. Widrow and S. D. Stearns, Adaptive Signal Processing. Upper Saddle River, NJ: Prentice-Hall, 1985. [4] A. D. C. Chan, “Multi-expert automatic speech recognition system using myoelectric signals,” Ph.D. dissertation, Univ. New Brunswick, Fredericton, NB, Canada, 2002. [5] M. S. Morse and E. M. O’Brien, “Research summary of a scheme to ascertain the availability of speech information in the myoelectric signals of neck and head muscles using surface electrodes,” Comput. Biol. Med., vol. 16, no. 6, pp. 399–410, 1986. [6] M. S. Morse, S. H. Day, B. Trull, and H. Morse, “Use of myoelectric signals to recognize speech,” in Proc. Annu. Int. Conf. the IEEE Engineering in Medicine and Biology Society—Images of the Twenty-First Century., Nov. 9–12, 1989, vol. 6, pp. 793–1794. [7] M. S. Morse, Y. N. Gopalan, and M. Wright, “Speech recognition using myoelectric signals with neural network,” in Proc. Annu. Int. Conf. IEEE Engineering in Medicine and Biology Society, 1991, vol. 13, no. 4, pp. 1877–1878.
699
[8] N. Sugie and K. Tsunoda, “A Speech prosthesis employing a speech synthesizer,” IEEE Trans. Biomed. Eng., vol. BME-32, no. 7, pp. 485–490, Jul. 1985. [9] H. Manabe, A. Hiraiwa, and A. Sugimura, “Unvoiced speech recognition using EMG—mime speech recognition,” in Proc. Conf. Human Factors in Computing Systems, Ft. Lauderdale, FL, 2003, pp. 794–795. [10] J. V. Basmajian and C. J. DeLuca, Muscles Alive: Their Functions Revealed by EMG, 5th ed. Baltimore, MD: Williams & Wilkins, 1985. [11] G. Aversano, A. Esposito, and M. Marinaro et al., “A new text-independent method for phoneme segmentation,” in Proc. 44th IEEE 2001 Midwest Symp. Circuits and Systems, 2001, vol. 2, pp. 516–519. [12] J. Kominek, C. L. Bennett, and A. W. Black, “Evaluating and correcting phoneme segmentation for unit selection synthesis,” in Proc. Eurospeech 2003, Geneva, Sep. 1–4, 2003, pp. 313–316. [13] O. Cappé, H2M: A Set of MATLAB/OCTAVE Functions for the EM Estimation of Mixtures and Hidden Markov Models ENST dpt. TSI/ LTCI (CNRS-URA 820), Aug. 2001. [14] L. R. Rabiner, “A tutorial on hidden markov models and selected applications in speech recognition,” Proc. IEEE, vol. 77, no. 2, Feb. 1989. [15] P. Roach, English Phonetics and Phonology: A Practical Course. Cambridge, U.K.: Cambridge Univ. Press, 1983. [16] P. F. Vint, S. P. McLean, and G. M. Harron, “Electromechanical delay in isometric actions initiated from nonresting levels,” Med. Sci. Sports Exer., vol. 33, no. 6, pp. 978–983, June 2001. [17] J. C. Junqua, “The lombard reflex and its role on human listeners and automatic speech recognizers,” J. Acoust. Soc. Am., vol. 85, no. 2, pp. 849–900, 1989. Erik J. Scheme received the B.Sc. and M.Sc. degrees in electrical engineering from the University of New Brunswick (UNB), Fredericton, NB, Canada, in 2003 and 2005, respectively. He is currently a project engineer at the Institute of Biomedical Engineering at UNB. His research interests include biological signal processing, speech and speaker recognition and the clinical application of research.
Bernard Hudgins (M’97–SM’01) received the Ph.D. degree from the University of New Brunswick (UNB), Fredericton, NB, Canada, in 1991. He is currently the Director of the Institute of Biomedical Engineering and Professor of Electrical and Computer Engineering at the same university. His primary research interests are in the areas of myoelectric signal processing for the control of artificial limbs and rehabilitation engineering. He recently spent two years on a NATO workgroup assessing alternative control technologies for cockpit applications Dr. Hudgins was the recipient of a Whitaker Foundation Young Investigator Award. He has been Region 7 (Canada) representative on the IEEE EMBS Advisory Committee, and EMBS VP for Publications and Technical Activities.
Philip A. Parker (S’70–M’77–SM’86) received the B.Sc. degree in electrical engineering from the University of New Brunswick (UNB), Fredericton, NB, Canada, in 1964, the M.Sc. degree from the University of St. Andrews, St. Andrews, U.K., in 1966, and the Ph.D. degree from UNB in 1975. In 1966, he joined the National Research Council of Canada as a Communications Officer and the following year he joined the Institute of Biomedical Engineering, UNB, as a Research Associate. In 1976, he was appointed to the Department of Electrical Engineering, UNB, and currently holds the rank of Professor Emeritus in that department. He is also a member of the Institute of Biomedical Engineering, UNB. His research interests are primarily in the area of biological signal processing.
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 4, APRIL 2007
Myoelectric Signal Classification for Phoneme-Based Speech Recognition Erik J. Scheme, Bernard Hudgins*, Senior Member, IEEE, and Phillip A. Parker, Senior Member, IEEE
Abstract—Traditional acoustic speech recognition accuracies have been shown to deteriorate in highly noisy environments. A secondary information source is exploited using surface myoelectric signals (MES) collected from facial articulatory muscles during speech. Words are classified at the phoneme level using a hidden Markov model (HMM) classifier. Acoustic and MES data was collected while the words “zero” through “nine” were spoken. An acoustic expert classified the 18 formative phonemes in low noise levels [signal-to-noise ratio (SNR) of 17.5 dB] with an accuracy of 99%, but deteriorated to approximately 38% under simulations with SNR approaching 0 dB. A fused acoustic-myoelectric multiexpert system, without knowledge of SNR, improved on acoustic classification results at all noise levels. A multiexpert system, incorporating SNR information, obtained accuracies of 99% at low noise levels while maintaining accuracies above 94% during low SNR (0 dB) simulations. Results improve on previous full word MES speech recognition accuracies by almost 10%.
MES (which is generated by muscles during contraction) is immune to acoustic noise, data collected from facial muscles during speech could be used to augment conventional acoustic recognition. This is particularly convenient for fighter pilots as electrodes can be placed inside the oxygen mask. MES classification is also a good complement to acoustic speech recognition (ASR) because sounds that are acoustically similar are often formed using reasonably different musculatures. Herein, Chan’s approach is expanded to classify words at the phoneme level. This provides a more detailed analysis by identifying the phonemic makeup of each word. Phoneme based recognition also allows for more rapid expansion of the recognizable vocabulary.
Index Terms—EMG, HMM, multi-expert systems, myoelectric signal, speech recognition.
II. BACKGROUND
I. INTRODUCTION
S
PEECH recognition has become a widely researched field covering tasks such as autonomous transcription on a home computer to advanced military and security applications. One proposed function is the control of secondary tasks in military fighter jets. Due to the complexity of their instrument panels and controls, pilots are forced to place a large portion of their focus downwards. It is desirable to implement a speech recognition system that would alleviate some of this dependence and enable more concentration to be shifted towards “heads-up” flying. However in order for speech recognition to be a usable tool in such a critical function accuracy rates must be exceptional. Due to the noisy environment in such aircraft, conventional speech recognition is limited. Although much research has been done to improve acoustic results under noisy conditions [1]–[3], it is beneficial to consider alternate approaches. Chan [4] proposed a multiexpert system using MES classification combined with acoustic recognition. Because the
Manuscript received November 25, 2005; revised September 16, 2006. This work was supported in part by the NSERC under Discovery Grant 171368-03 and Discovery Grant 4445-04 and in part by the New Brunswick Innovation Fund RAI program. Asterisk indicates corresponding author. E. J. Scheme is with the Institute of Biomedical Engineering and the Department of Electrical and Computer Engineering, University of New Brunswick, Fredericton, NB E3B 5A3, Canada. *B. Hudgins is with the Institute of Biomedical Engineering, University of New Brunswick, PO Box 4400, Fredericton, NB E3B 5A3, Canada (e-mail: [email protected]). P. A. Parker is with the Institute of Biomedical Engineering, University of New Brunswick, Fredericton, NB E3B 5A3, Canada. Digital Object Identifier 10.1109/TBME.2006.889175
The first appearances of MES based voice or speech prostheses occurred in the mid 1980s. Morse and O’Brien [5] used four channels of MES data (three recorded from the neck and one from the forehead) to obtain a 97% classification rate for a 2-word vocabulary. Later, Morse et al. reported accuracies of 57% [6] and 60% [7] for 10-word vocabularies. Sugie et al. [8] suggested the use of a voice prosthesis for discriminating between the five Japanese vowels. Using an automaton, they were able to achieve an average classification rate of approximately 60% using three channels of MES recorded from sites around the mouth. Chan et al. [4] proposed a method of speech recognition using a multiexpert system combining both acoustic and MES recognition. In a preliminary study using 5 channels of facial MES data, Chan was able to obtain an accuracy of approximately 85% and 87% when classifying the English words “zero” through “nine” for two subjects. When combined with an acoustic expert, this system proved to greatly improve recognition rates in noisy environments when compared to a purely acoustic classifier. In 2003, Manabe et al. [9] used MES data recorded from three facial muscles during unvoiced speech to perform what they called “mimed speech recognition.” Three subjects mouthed the five Japanese vowels for periods of 10 s, and 400 ms portions were later extracted to form training and test sets. Using a three-layer neural network, Manabe obtained an average classification rate of over 94%. The vowels were not taken contextually from speech and were mouthed for extended periods, effectively eliminating any transient or coarticulatory effects which may have occurred in conversational speech. Although research in this field has been relatively limited, results have shown that significant information exists in the MES
0018-9294/$25.00 © 2007 IEEE
SCHEME et al.: MES CLASSIFICATION FOR PHONEME-BASED SPEECH RECOGNITION
695
IV. SEGMENTATION
Fig. 1. Electrode placement.
of articulatory muscles. It has been shown that speech information is present at the word level and from separate vocalizations of vowels. Herein, we confirm the presence of phonemic data in the MES recorded from spoken words. III. METHOD As in Chan’s study, five channels of MES data were recorded using Ag-AgCl Duotrode bipolar electrode pairs placed over five articulatory muscles of the face: the levator anguli oris, the zygomaticus, the platysma, the depressor anguli oris, and the anterior belly of the digastricus. Unlike in Chan’s work, electrodes were not embedded in a pilot’s oxygen mask because the mask severely degraded the quality of acoustic speech needed for phonemic segmentation. Instead, electrode pairs were placed directly on the speakers face (Fig. 1), and taped as required to minimize any extraneous movement during speech. The subject’s skin was cleaned using alcohol swabs to remove any dirt or makeup, and electrode gel was added to the electrodes to minimize impedance at the skin-electrode interface. An Ag-AgCl Red-Dot electrode (3M, 2259) was also placed on the back of the neck to provide a common reference. All signals were then sampled at 5 kHz using a simultaneous sample and hold device (ComputerBoards, CIO SSH16) and a 16-bit analog to digital board (Measurement Computing, PCIMDAS1602/16). Four English speaking subjects (3 male, 1 female) each performed 9 sets of 5 repetitions of the words “zero” through “nine.” Subjects were encouraged to rest between sets and the words were presented in random order to minimize anticipatory articulation. It was found, in early trials, that when a subject knew the next word to be spoken, there was a tendency to preposition the facial muscles earlier than in natural speech. All training and testing data were collected during one session due to subject availability. A multisession experiment would require electrode placement to be carefully reproduced. However, the classifier exploits interchannel patterns and it is, therefore, believed that reasonable amounts of intersession variability are tolerable and trained models can be used across sessions. A mask embedded with electrodes, such as that used by Chan [4] but causing less acoustic distortion, could sufficiently minimize electrode placement variability.
In order to train a phoneme classifier, it is necessary to first have speech that is segmented into phonemes. However, there is currently no published information about the phonemic segmentation of MES. Also, considering that the MES of two individuals is distinct [10] (making it impossible to form a general MES phoneme model across subjects) it quickly becomes apparent that directly segmenting the MES is a challenge. Morse and O’Brien [5] showed that a slight delay exists between the onset of MES from facial muscles and the appearance of its associated acoustic speech. This is due to the fact that the muscles in the vocal tract must be contracted prior to the generation of sound. A slight delay is also apparent due to the time required for sound to propagate to a microphone. It was, therefore, decided to segment the acoustic data, apply a negative time shift and relate the corresponding segments/labels to the associated MES. Current automatic phoneme segmentation algorithms [11] correctly segment approximately 74% of phonemes without over-segmenting. The incorrectly labeled data is not only skewed, but false phonemes can be inserted, and correct phonemes disregarded. Therefore, even when an automatic labeler is implemented, a manual verification of each and every result is required [12]. Both the training and test data for this experiment was, therefore, manually segmented. Segmentation was facilitated by using custom Matlab software that allowed the user to view spectrographic and formant information as well as use audio cues for labelling phonemes. During initial segmentation trials, phonemes were chosen to be continuous within a word, starting immediately after the preceding phoneme ended. It was later found that results could be improved when segments were labelled more stringently to avoid incorporating coarticulatory data. This effect exists during the natural flow of speech at the time when two phonemes are strung together, thereby changing their characteristics. It is possible that coarticulation effects may even be exaggerated in MES due to anticipatory muscle contractions which occur while prepositioning muscles for articulation. The acoustic speech data was, therefore, manually re-segmented with focus on selecting only those parts caused purely by the desired phoneme (Fig. 2). Again, time-shifted versions of the resulting labels were applied to the MES data. Sixty percent of each speaker’s data set was used for training, while the remaining 40% was used for testing. It was found that increasing the number of repetitions in training beyond 27 (60% of 45) produced diminished returns in classification accuracy. The test set for each speaker, therefore, consisted of 18 repetitions of each word. Once the training words were chosen, the associated phonemes were sorted into input training vectors for each HMM phoneme model. The phonemic representation of the words as used herein is shown in Table I. Rather than phonemes, as used herein, some systems classify triphones which incorporate information from the phonemes preceding and following a particular phone. This has been shown to increase performance of recognition systems, but at the cost of greatly increasing the size of the collection of combinations classifiable segments. There are
696
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 4, APRIL 2007
TABLE II CLASSIFICATION RESULTS FOR ACOUSTIC DATA
Fig. 2. Acoustic segmentation of the word “zero” with negative time shift applied to MES data.
TABLE I PHONEMIC COMPOSITION OF TEST DATA
of 3 phonemes. While the English language limits the number of allowable combinations, there are still thousands of triphones compared to the 44 phonemes. Many research projects use a selection of words chosen to minimize the number of spoken triphones. In the 18 phoneme vocabulary used in this research, there are 30 different triphones. V. CLASSIFICATION The HMM classifier used during this research was developed using the H2M toolbox for Matlab [13]. Five state, left-to-right HMM were trained with single Gaussian mixture observation densities using the Forward-Backward algorithm [14] for each phoneme. Because each person’s MES is unique, speaker dependent models were required. Features for the acoustic classifier consisted of the root-mean-square value of each frame and the first 12 Mel-cepstral frequency coefficients (MFCC). Because of their close approximation to the human perception of
sound, MFCC have become widely used in ASR. However, because the Mel-scale warping is tailored to the human perception of acoustic speech, it would initially appear that this warping of the frequency spectrum might not apply to MES. However, Chan [4] found that MFCC performed better than other feature sets when classifying facial MES recorded during speech. The same feature set was, therefore, used to represent each of the five channels of data for the MES classifier. While Chan [4] used frames of 64 ms, window lengths herein were restricted due to the small size of certain phonemic segments. It was necessary to constrain the frame length and increment in order to obtain a reasonable number of frames when representing short phonemes. Classification performance suffered when greater window lengths were chosen because information from adjoining phonemes became included in the frame. While these parameters were investigated separately for acoustic and MES data (with frame lengths from 5 ms to 100 ms), the configuration that provided the best classification accuracy for both consisted of frame lengths of 12.8 ms, overlapped by 1.6 ms. Phoneme classification decisions were made using the maximum likelihood output of the Viterbi algorithm [14]. For a given segment of speech (a word), the phoneme classification results were then compared to those in a dictionary and the segment was assigned a score based on the number of matching phonemes. The dictionary consisted of a list of the test words (“zero” through “nine”) along with their phonemic breakdowns. The dictionary word which corresponded to the highest score was then selected as the recognized word. It is possible for two words to return the same maximum score. For example, the phonemes “f”-“i”-“r” would return an equal score for the words “four” (“f”-“oh”-“r”) and “five” (“f”“i”-“v”). In this case, a likelihood score was determined by summing the likelihoods associated with each word. This was done by summing the original likelihoods for each phoneme, as obtained from the classification process. The maximum likelihood would then be selected and the associated word chosen. VI. RESULTS The acoustic data was classified using the discussed HMM classifier. The results of the classification are shown in Table II. While the average phoneme classification accuracy is slightly over 91%, the overall word accuracy is over 99%. This illustrates one of the benefits of using a segmental approach to speech classification. The combination of multiple phonemes
SCHEME et al.: MES CLASSIFICATION FOR PHONEME-BASED SPEECH RECOGNITION
697
TABLE III CLASSIFICATION RESULTS FOR MES DATA
Fig. 3. Effect of offset between acoustic and MES segmentation boundaries on overall classification accuracy.
into one word (following a structured grammatical format) allows a certain level of forgiveness. If a phoneme within a word is misclassified, there may be sufficient information provided by the other phonemes to correctly identify the word. Because there is a delay between the onset of MES activity and the appearance of acoustic speech [5], a time shifted variation of the acoustic segmentation was used to segment the MES data. The choice of offset is crucial because it is necessary to have the exact MES information that corresponds to the acoustic phoneme being spoken. Chan [4] applied a shift value of 500 ms to make sure that all MES data was included in the overall time window representing a spoken word. However, for phoneme based recognition it is necessary to further scrutinize this delay. During speech, natural silences exist between words, allowing some room for error in the location of a selection window. However, because phonemes within a word are typically strung together without pause, the window must (optimally) envelope the desired phoneme without encompassing any extra speech. To test this, classification accuracy was measured as a function of offset/delay. The average effect is shown in Fig. 3. It can be seen that a distinct peak exists around the 100 ms mark. This implies that the optimal correlation between the acoustic and MES data occurs at a delay of 100 ms. This is consistent with the literature for typical electro-mechanical delays in muscle [16]. Having selected the optimal offset, classification of the MES was performed. Table III shows the obtained results. These results show an almost 10% improvement over results obtained by Chan [4] in a study of the effect of temporal misalignment on MES-based speech recognition. Chan used full word HMM classification to attain an accuracy of approximately 85% using the same 10-word vocabulary.
In order to better demonstrate the merits of using MES speech recognition, it is beneficial to investigate its performance in acoustically noisy environments. Varying levels of white Gaussian noise were added to the collected acoustic data so that the SNR (calculated using a ratio of signal power) resulting from the most severely degraded word sample, on a per speaker basis, was approximately 0 dB. This was determined empirically so that acoustic phoneme recognition rates dropped to approximately a priori at the highest noise level. The other levels of noise were distributed between clean (the unaltered collected data was designated as the “clean” class) and the 0 dB SNR. HMM were trained using the unaltered data and classification was performed on all levels of the noise-corrupted data. No acoustic noise filtering or cancellation was implemented in order to better illustrate the effect of noise on acoustic classification rates. It can be seen in Fig. 4 that the acoustic classification accuracy falls to 8% at 0 dB, approximately the a priori rate of 6% for 18 phonemes. For the purpose of this investigation, the Lombard effect has been assumed to be negligible. The Lombard effect is the change in the way a person speaks when in acoustically noisy conditions [17]. It has been suggested that changes in power, shifting of formant frequencies, and changes in formant band width (among others) can occur, however Chan [4] found that the effect did not appreciably change facial MES at the noise levels used herein. Because, under this assumption, the MES is essentially immune to acoustic noise, the collected MES data was assumed to be constant across all levels of acoustic noise. This is depicted in Fig. 4 as a horizontal line for MES classification rates across varying acoustic noise levels. Note that the MES classification rate is inferior to the acoustic rate for only low levels of noise. When higher levels of noise are introduced, the MES becomes a better source of information for phoneme recognition. Because the classifiers perform differently in varying noise levels, it is desirable to exploit their relative strengths. A multiexpert system combines the opinions of two or more classifiers, yielding an overall classification decision which is a function of the individual verdicts of its formative experts. Chan [4] describes several ways of forming multiexpert systems, incorporating various levels of abstractness (majority vote, Bayesian belief, Dempster-Shafer, etc). In particular, Chan introduces a plausibility method of combining experts using mathematical theory of evidence. The method assigns weights to the formative classifiers based
698
IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 4, APRIL 2007
TABLE IV AVERAGE PHONEME ACCURACY (%) VERSUS SNR
TABLE V AVERAGE WORD ACCURACY (%) VERSUS SNR
Fig. 4. Phoneme (top) and word (bottom) classification accuracies of the acoustic, MES, and multiexpert systems (with and without knowledge of acoustic SNR) versus acoustic SNR.
on their ability to discriminate information. The algorithm is particularly beneficial in applications where the SNR cannot be measured, because it is able to assign beliefs based on the distribution of likelihoods, without empirical knowledge of performance. A plausibility based multiexpert system was implemented, combining the likelihoods of the acoustic and MES classifiers. The phoneme corresponding to the maximum resulting likelihood was selected as the multiexpert result. Fig. 4 shows that the fused system [denoted as “Multi-Expert (No SNR)”] outperforms the acoustic classifier alone at all noise levels. This confirms that the MES classifier is able to discern information orthogonal to that of the acoustic classifier, and can be used to improve accuracy even in low noise environments. At high noise levels, however, the classification accuracy of the multiexpert system drops below that of the MES classifier. An optimal fusion system would produce accuracies equal to or greater than that of the most accurate classifier at all noise levels. When the acoustic SNR ratio is known, it is possible to use heuristic information about classifier accuracies in the presence of noise to determine optimal weightings for a multiexpert combination equation. The experts were, therefore, combined using a weighted sum of their outputs, where the weights were a function of each classifier’s accuracy (the ensemble mean across all subjects) in varying levels of acoustic SNR. Matlab’s curve fitting toolbox was used to approximate a continuous weighting curve for the
acoustic classifier (shown in (1)). Because the phoneme accuracy of the MES classifier is independent of acoustic noise, its weight was a constant,
where (1) , was calculated by The multiexpert likelihood, weighting the acoustic classifier likelihood, , and adding it to a weighted version of the MES classifier likelihood, [shown in (2)]. The phoneme with the highest likelihood was then selected in the same way it was for each individual classifier (2) While this system requires knowledge of the SNR, it can be seen in Fig. 4 that it matches or outperforms both the acoustic and MES classifier at all noise levels. These results are also shown below in Tables IV and V. VII. CONCLUSION Phoneme based acoustic and MES speech recognition systems were presented. MES were extracted from 5 articulatory muscles in the face as a secondary source of speech information. The digits from “zero” through “nine” were used as the training
SCHEME et al.: MES CLASSIFICATION FOR PHONEME-BASED SPEECH RECOGNITION
and test vocabulary. Manual segmentation was performed in order to demonstrate the accuracy of the phoneme recognition without performance degradation due to errors in automatic segmentation. The acoustic system outperformed the MES system at low noise levels (99% to 95%), however its performance significantly deteriorated at lower SNRs, while the acoustic noise immune MES system upheld its accuracy. It was shown that a multiexpert system combining the opinions of both classifiers, without knowledge of SNR, outperformed the acoustic classifier at all noise levels. A multiexpert system, incorporating information about SNR, outperformed both classifiers at all noise levels, obtaining word accuracies above 99% at low noise levels and above 94% at noise levels approaching 0 dB. In addition, MES classification at the phoneme level yielded accuracies almost 10% better then those obtained by Chan [4] using full word recognition for the same vocabulary. MES phoneme recognition improves upon previous wordbased MES systems by enabling the addition of new words without further training. If an additional word were to be added to a word-based system, a new model would first need to be trained and tested. Using the method discussed herein, any word comprised of the current phonemes (18 have been trained out of the approximately 44 in the English language [15]) can be added to the dictionary without additional training. A system containing all 44 phonemes would be able to be expanded to classify any word in the English language by simply adding that word to the dictionary. A maximum of 44 HMM would, therefore, be necessary, whereas the number of HMM would increase at a 1:1 ratio with the number of words added to a full word recognition system. Future research will investigate the expansion of the phoneme database, as well as the feasibility and accuracy of real-time acoustic-MES speech recognition using automatic segmentation of speech. Alternate sources of noise, such as electromagnetic and effects of gravity on speech will also be investigated.
REFERENCES [1] B. Widrow, “Adaptive noise cancelling: principles and applications,” Proc. IEEE, vol. 63, no. 12, pp. 1692–1716, Dec. 1975. [2] M. R. Sambur, “Adaptive noise canceling for speech signals,” IEEE Trans. Acoust., Speech, Signal Process., vol. ASSP-26, no. 5, pp. 419–423, Oct. 1978. [3] B. Widrow and S. D. Stearns, Adaptive Signal Processing. Upper Saddle River, NJ: Prentice-Hall, 1985. [4] A. D. C. Chan, “Multi-expert automatic speech recognition system using myoelectric signals,” Ph.D. dissertation, Univ. New Brunswick, Fredericton, NB, Canada, 2002. [5] M. S. Morse and E. M. O’Brien, “Research summary of a scheme to ascertain the availability of speech information in the myoelectric signals of neck and head muscles using surface electrodes,” Comput. Biol. Med., vol. 16, no. 6, pp. 399–410, 1986. [6] M. S. Morse, S. H. Day, B. Trull, and H. Morse, “Use of myoelectric signals to recognize speech,” in Proc. Annu. Int. Conf. the IEEE Engineering in Medicine and Biology Society—Images of the Twenty-First Century., Nov. 9–12, 1989, vol. 6, pp. 793–1794. [7] M. S. Morse, Y. N. Gopalan, and M. Wright, “Speech recognition using myoelectric signals with neural network,” in Proc. Annu. Int. Conf. IEEE Engineering in Medicine and Biology Society, 1991, vol. 13, no. 4, pp. 1877–1878.
699
[8] N. Sugie and K. Tsunoda, “A Speech prosthesis employing a speech synthesizer,” IEEE Trans. Biomed. Eng., vol. BME-32, no. 7, pp. 485–490, Jul. 1985. [9] H. Manabe, A. Hiraiwa, and A. Sugimura, “Unvoiced speech recognition using EMG—mime speech recognition,” in Proc. Conf. Human Factors in Computing Systems, Ft. Lauderdale, FL, 2003, pp. 794–795. [10] J. V. Basmajian and C. J. DeLuca, Muscles Alive: Their Functions Revealed by EMG, 5th ed. Baltimore, MD: Williams & Wilkins, 1985. [11] G. Aversano, A. Esposito, and M. Marinaro et al., “A new text-independent method for phoneme segmentation,” in Proc. 44th IEEE 2001 Midwest Symp. Circuits and Systems, 2001, vol. 2, pp. 516–519. [12] J. Kominek, C. L. Bennett, and A. W. Black, “Evaluating and correcting phoneme segmentation for unit selection synthesis,” in Proc. Eurospeech 2003, Geneva, Sep. 1–4, 2003, pp. 313–316. [13] O. Cappé, H2M: A Set of MATLAB/OCTAVE Functions for the EM Estimation of Mixtures and Hidden Markov Models ENST dpt. TSI/ LTCI (CNRS-URA 820), Aug. 2001. [14] L. R. Rabiner, “A tutorial on hidden markov models and selected applications in speech recognition,” Proc. IEEE, vol. 77, no. 2, Feb. 1989. [15] P. Roach, English Phonetics and Phonology: A Practical Course. Cambridge, U.K.: Cambridge Univ. Press, 1983. [16] P. F. Vint, S. P. McLean, and G. M. Harron, “Electromechanical delay in isometric actions initiated from nonresting levels,” Med. Sci. Sports Exer., vol. 33, no. 6, pp. 978–983, June 2001. [17] J. C. Junqua, “The lombard reflex and its role on human listeners and automatic speech recognizers,” J. Acoust. Soc. Am., vol. 85, no. 2, pp. 849–900, 1989. Erik J. Scheme received the B.Sc. and M.Sc. degrees in electrical engineering from the University of New Brunswick (UNB), Fredericton, NB, Canada, in 2003 and 2005, respectively. He is currently a project engineer at the Institute of Biomedical Engineering at UNB. His research interests include biological signal processing, speech and speaker recognition and the clinical application of research.
Bernard Hudgins (M’97–SM’01) received the Ph.D. degree from the University of New Brunswick (UNB), Fredericton, NB, Canada, in 1991. He is currently the Director of the Institute of Biomedical Engineering and Professor of Electrical and Computer Engineering at the same university. His primary research interests are in the areas of myoelectric signal processing for the control of artificial limbs and rehabilitation engineering. He recently spent two years on a NATO workgroup assessing alternative control technologies for cockpit applications Dr. Hudgins was the recipient of a Whitaker Foundation Young Investigator Award. He has been Region 7 (Canada) representative on the IEEE EMBS Advisory Committee, and EMBS VP for Publications and Technical Activities.
Philip A. Parker (S’70–M’77–SM’86) received the B.Sc. degree in electrical engineering from the University of New Brunswick (UNB), Fredericton, NB, Canada, in 1964, the M.Sc. degree from the University of St. Andrews, St. Andrews, U.K., in 1966, and the Ph.D. degree from UNB in 1975. In 1966, he joined the National Research Council of Canada as a Communications Officer and the following year he joined the Institute of Biomedical Engineering, UNB, as a Research Associate. In 1976, he was appointed to the Department of Electrical Engineering, UNB, and currently holds the rank of Professor Emeritus in that department. He is also a member of the Institute of Biomedical Engineering, UNB. His research interests are primarily in the area of biological signal processing.
