Generation and Analysis of Vowel Polygons

Czech vowels in fluent speech and their parameters can be found by The Generation and Analysis of Vowel Polygons software. On the base of found vowels formant values, vowel polygons can be created and further analyzed as a unit in relevant formant planes. Phoneme recognition is based on pattern similarity using Mahalanobis distance and Feed-forward Neural Network, and it is expanded by time duration control of found phoneme [1] and voluntary data reduction achieving more pure results. Output data are not only presented in graphic form, but they are also stored statistically in *.xls files. Found vowels are saved to *.wav sound files. The presented software can be helpful in e.g. actual state of speaker recognition or its identification [2], [3]. Fig. 1: The processing of chosen speech record (left), the presentation of achieved results and the analysis of relevant vowel polygon (right). [1] STANĚK, M., POLÁK, L., Algorithms for vowel recognition in fluent speech based on formant position. Proc. Int. Conf. on Telecommunications and Signal Processing, Rome, Italy, July 2013. [2] STANĚK, M. Comparison of speaker individuality in areas of plane formant spaces. Electronics Letters, under review. [3] STANĚK, M., SIGMUND, M., Finding the Most Suitable Property of Created Normalization Vector for Speaker Recognition. Radioengineering, under review.