Article 9215

Title of the article

METHOD FOR OBTAINING VECTORS OF ACOUSTIC CHARACTERISTICS FOR PHRASE SEQUENCE RECOGNITION IN CONDITIONS OF NOISE INTERFERENCE

Authors

Stas Tambi Takhsinovich, Leading programming engineer, scientific production association Russian innovative technologies (building 1, 14 Ozernaya street, Tver, Russia), ttstas@npo-rit.ru
Shcherbakov Mikhail Aleksandrovich, Doctor of engineering sciences, professor, head of sub-department of automation and remote control, Penza State University (40 Krasnaya street, Penza, Russia), avitel@pnzgu.ru
Sazonov Vladimir Vasil'evich, Candidate of engineering sciences, associate professor, sub-department of automation and remote control, Penza State University (40 Krasnaya street, Penza, Russia), sazonov@inbox.ru

Index UDK

621.391

Abstract

Background. Improvement of the accuracy of speech recognition under noise interference is an important task, because there are applications of speech recognition systems, where, despite the growing relevance, there is a problem of high recognition accuracy achievement in the presence of various noise impacts. The research object is on-board systems of automatic speech recognition. The research subject is the improvement of speech recognition probability under noise interference. The work purpose is to develop methods and algorithms of the recognition accuracy improvement under noise interference.
Materials and methods. The research of speech recognition methods under noise interference was implemented using neural networks and hidden Markov models.
Results. The authors have developed a method of obtaining a vector of acoustic features based on the use of chalk-frequency cepstral coefficients. The basis of this method of training is formed by a new formula of linear single-layer neural networks (LONS), obtained through the use of two target functions. The first objective function – a function of the probability of the normal Gaussian multivariate distribution, the second objective function – a function of calculating the cepstral coefficients based on the use of LONS for computing the average spectral power.
Conclusions. The authors have suggested the method of obtaining a vector of acoustic features based on the use of chalk-frequency cepstral coefficients. The LONS learning algorithm, which is based on the formula of training obtained through the application of the two objective functions, enhanced the probability of speech recognition under noise interference. 

Key words

neural network, hidden Markov model, speech recognition, on-board equipment.

Download PDF
References

1. Englund C. Speech recognition in the JAS 39 Gripen aircraft -Adaptation to speech at dierent G-loads: Master Thesis in Speech Technology. Department of Speech, Music and Hearing, Royal Institute of Technology. Stockholm. 2004, 11th March, p. 1.
2. Vaseghi S. V. Digital signal processing and noise reduction, fourth edition. London: John Wily and Sons Ltd Publication, 2008, 466 p.
3. Dongsuk Y. Robust speech recognition using neural networks and hidden markov mod-els: Doctor of Philosophy.New Jersey:Graduate school–New Brunswick Rutgers,The State University of New Jersey,1999,18p.
4. Giampiero S. Mining speech sounds, machine learning methods for automatic speech recognition and analysis: doctoral thesis. Stockholm: KTH school of computer science and communication, 2006, 25 p.
5. Geppener V., Simonchik K. Sovremennye problemy neyroinformatiki [Modern problems of neural informatics]. 2006, no. 15, pp. 14–23.
6. Fil'tr s beskonechnoy impul'snoy kharakteristikoy [Filters with infinite impulse re-sponse]. Available at: https://ru.wikipedia. org/wiki/Fil'tr_s_beskonechnoy_impul'snoy_kharakteristikoy
7. Passel S., Norvig P. Iskusstvennyy intellekt – sovremennoy podkhod [Artificial intelli-gence – modern approach]. Moscow: Vil'yams, 2007, 777 p.
8. Rabiner L., Juang B.-H. Fundamentals of sreech recognition. New Jersey: Prentice Hall, 1993, 387 p.
9. Tikhomirov V. A., Stas T. T. Neobratimye protsessy v prirode i tekhnike: sb. Vserossiy-skoy konf. (Moskva, 26–28 yanvarya 2011 g.) [Irreversible processes in nature and tech-nology: proceedings of the All-Russian conference (Moscow, 26-28 January 2011)]. Moscow: Izd-vo MGTU im. Baumana, 2011, part II, no. 11, pp. 140–144.
10. Dzh T., Gonsales R. Printsipy raspoznavaniya obrazov [Concepts of image recogni-tion]. Moscow: Mir, 1978, 242 p.

 

Дата создания: 02.10.2015 15:15
Дата обновления: 05.10.2015 13:38