**8. Conclusions and future work**

In this chapter was presented a method to implement a high performance, real-time, restricted-vocabulary speech recognition system, combining a genetic algorithm and the Fisher's Linear Discriminant Ratio (FLDR) in its matrix formulation. A review of the concepts of variable and feature selection as well as feature generation was made; also were presented some concepts related with speech processing, like the LPC formulation and the DTW method for template matching.

One of the conceptual tools used here was the energy of the signal in certain sub-bands in the frequency domain; thanks to the Parseval's theorem, the same amounts of energy can be calculated in the time domain via a bank of digital filters, enabling thus a very fast way to apply the recognizer, since the process goes on at the same time as the occurrence of the word is exerted. Mainly, two experiments were shown, in Spanish and English, with male and female participants; in both cases high performance was attained, beyond 94% at the

Optimal Feature Generation with

Thesis 1993.

2001.

Genetic Algorithms and FLDR in a Restricted-Vocabulary Speech Recognition System 261

[11] L. D. Vignolo, H. L. Rufiner, D. H. Milone, and J. C. Goddard, Evolutionary Splines for

[12] T. Takiguchi, N. Miyake, H. Matsuda, and Y. Ariki, Voice and Noise Detection with

[13] S. Y. Suk and H. Kojima, Voice Activated Appliances for Severely Disabled Persons, in

[14] R. Cardin, Improved Learning Strategies for Small Vocabulary Automatic Speech

[15] Isabelle Guyon. (2011, june) Isabelle Guyon's home page. [on-line]. Hyperlink

[16] I. S. Oh, J. S. Lee, and B. R. Moon, Hybrid Genetic Algorithms for Feature Selection,

[17] P. R. Somol, J. Novovicova, and P. Pudil, Efficient Feature Subset Selection and Subset

[18] Y. Sun, S. Todorovic, and S. Goodison, Local-Learning-Based Feature Selection for

[19] J. Teixeria de Souza, R. A. Ferreira do Carmo, and G. Campos de Lima, On the

[20] X. Huang, A. Acero, and H. W. Hon, *Spoken Language Processing. A guide to Theory,* 

[21] A. Zacknich, *Principles of Adaptive Filters and Self-Learning Systems*, 1st ed. London,

[22] H. Sakoe and S. Chiba, Dynamic Programming Algorithm Optimization for Spoken

[23] P. J. Bigus and J. Bigus, *Constructing Intelligent Agents with Java. A Programmer´s Guide to* 

[24] K. S. Tang, K. F. Man, S. Kwong, and Q. He, Genetic Algorithms and their Applications, *IEEE Signal Processing Magazine*, pp. 22-37, November 1996. [25] S. M. Sait and A. Youssef, *Iterative Computer Algorithms with Applications in Engineering.*,

[26] S. M. Ahadi, H. Sheikhzadeh, R. L. Brennan, and G. H. Freeman, An Effective Front-

[27] M. P. G. Saon, G. Zweig, J. Huang, B. Kingsbury, and L. Mangu, Evolution of the

*Intelligence*, vol. 32, no. 9, pp. 1610-1626, September 2010.

Zhang, Ed. Vienna, Austria: In-Tech, 2010, cap. 9, pp. 158-171.

*Smarter Applications.*, 1st ed.: John Wiley & Sons, Inc., 2001.

1st ed. Los Alamitos, Cal., U.S.A.: IEEE Computer Society, 1999.

*on Advances in Signal Processing*, vol. 2011, 2011, pp. 1-15.

Vienna, Austria: In-Tech, 2008, cap. 29, pp. 527-538.

http://www.clopinet.com/ isabelle/.

Austria: InTech, 2010, cap. 4, pp. 75-98.

1424-1437, November 2004.

England: Springer, 2005.

26, no. 1, pp. 43-49, 1978.

band speech recognition.

Kroschel, Eds. Vienna, Austria: In-Tech, 2007, cap. 4, pp. 67-74.

Cepstral Filterbank Optimization in Phoneme Classsification, in *EURASIP Journal* 

AdaBoost, in *Robust Speech Recognition and Understanding*, M. Grimm and K.

*Speech Recognition, Technologies and Applications*, F. Mihelic and J. Zibert, Eds.

Recognition, McGill University, Montreal, Quebec, Canadá, Doctor of Philosophy

*IEEE Transactions on Pattern Analysis and Machine Intelligence*, vol. 26, no. 11, pp.

Size Optimization, in *Pattern Recognition Recent Advances*, A. Herout, Ed. Vienna,

High-Dimensional Data Analysis, *IEEE Transactions of Pattern Analysis and Machine* 

Combination of Feature Selection and Instance Selection, in *Machine Learning*, Y.

*Algorithm, and System Development*, 1st ed. New Jersey, U.S.A.: Prentice Hall PTR,

Word Recognition , *IEEE Transactions on Acoustics, Speech, and Signal Processing*, vol.

End for Automatic Speech Recognition, in *2003 International Conference on Electronics, Circuits and Systems*, Sharah, United Arab Emirates, 2003, pp. 1-4, Sub-

Performance of Automatic Speech Recogntion Algorithms in Transcribing

worst case. Compared to a typical implementation with backpropagation neural networks and cepstral coefficients, this approach was at least 10% more effective in near real-life application.

The genetic algorithm consistently maximized the criteria of inter-class separability and intra-class compactness, under different conditions of population, probability of mutation, etc., and also varying the restriction set. It is remarkable and worth to mention that the genetic algorithm didn't gave the best result in its first execution, which means that the execution must be repeated to achieve good results; increasing the population and manipulating the restriction set demonstrated that it is possible to obtain a variety of different outcomes, so it is important to experiment carefully.

The future work will consist of developing the voice/unvoiced detection in noisy environments, investigate an adapt more features that can be easily computed in time and with a dual in frequency, start working on the non-dependant speaker approach, making use of more robust classifiers, and finally, increase the vocabulary size, although still restricted to a specific semantic field, like in [35]. Another interesting venue is the one in which the user aging process is taken into account by the speech recognition system [36].
