Adam Dustor

Biometric Voice Identification Based on Fuzzy Kernel Classifier

This paper presents research on automatic speaker identification based on structural risk minimization and kernel functions. New approach, known as a Fuzzy Kernel Ho-Kashyap classifier FKHK, to speaker identification was applied. Instead of the most popular kernel functions like gaussian or polynomial, data dependent kernel matrix which may be interpreted in terms of linguistic values from the premises of if-then rules was applied. Classifier was tested on polish speech corpora ROBOT and obtained results were discussed.

Automatic Speech Segmentation for Automatic Speech Translation

The article presents selected, effective speech signal processing algorithms and their use in order to improve the automatic speech translation. Automatic speech translation uses natural language processing techniques implemented using algorithms of automatic speech recognition, speaker recognition, automatic text translation and text-to-speech synthesis. It is very possible to improve the process of automatic speech translation by using effective algorithms for automatic segmentation of speech signals based on speaker recognition and language recognition.

Speaker Verification Based on Fuzzy Classifier

This article presents a new type of a fuzzy classifier applied to speaker verification. Obtained decision function is nonlinear. Achieved verification accuracy is compared with classical techniques like Gaussian mixture models and vector quantization. Polish speech corpora ROBOT was applied as a training and testing set. Obtained results are discussed.

Speaker recognition system with good generalization properties

This paper presents speaker recognition system possessing very good generalization properties. Relatively low equal error rate for speaker verification and high identification rate for identification are achieved for very short training and testing sequences. This behaviour is achieved for the kernel modification of a classic Ho-Kashyap linear classifier. Achieved results for the new approach are compared with results for the classic GMM and VQ techniques. Speech of a moderately good quality from the Polish speech corpus was used for development of recognition system.

Influence of Feature Dimensionality and Model Complexity on Speaker Verification Performance

This paper provides description of a text dependent speaker recognition system based on vector quantization approach. The scope of this paper is to check influence of feature dimensionality and the complexity of the speaker model on verification process. Provided results show that MFCC features yield the lowest possible verification errors among all tested parameters. Although dimensionality of feature vectors is important, there is no need to increase it above some level as the improvement in verification performance is relatively low and computational complexity increases. Far more important than dimensionality is complexity of the speaker model.

Voice verification based on nonlinear Ho-Kashyap classifier

This article presents enhancement of a linear Ho-Kashyap classifier based on kernel functions applied to speaker verification. Obtained decision function is nonlinear. Achieved verification accuracy is compared with classical techniques like Gaussian mixture models and vector quantization. Polish speech corpora ROBOT was applied as a training and testing set. Obtained results are discussed.

Speech Recognition Based on Open Source Speech Processing Software

Creating of speech recognition application requires advanced speech processing techniques realized by specialized speech processing software. It is very possible to improve the speech recognition research by using frameworks based on open source speech processing software. The article presents the possibility of using open source speech processing software to construct own speech recognition application.

Influence of Corpus Size on Speaker Verification

The scope of this paper is to check influence of the size of the speech corpus on the speaker recognition performance. Obtained results for TIMIT corpus are compared with results obtained for smaller database ROBOT. Additionally influence of feature dimensionality and size of the speaker model was tested. Achieved results show that the best results can be obtained for MFCC features. The lowest EER for larger TIMIT database are 4 times worse than the best result for ROBOT corpus which confirms that biometric systems should be tested on as large data sets as possible to assure that achieved error rates are statistically significant.

Speaker Verification Performance Evaluation Based on Open Source Speech Processing Software and TIMIT Speech Corpus

Creating of speaker recognition application requires advanced speech processing techniques realized by specialized speech processing software. It is very possible to improve the speaker recognition research by using speech processing platform based on open source software. The article presents the example of using open source speech processing software to perform speaker verification experiments designed to test various speaker recognition models based on different scenarios. Speaker verification efficiency was evaluated for each scenario using TIMIT speech corpus distributed by Linguistic Data Consortium. The experiment results allowed to compare and select the best scenario to build speaker model for speaker verification application.

Matlab Based Closed Set Speaker Recognition

Abstract This paper presents the fundamental part of all automatic speaker recognition systems (ASR) which is namely pattern recognition used to measure similarity between speaker model stored in a system and parameters extracted from the test utterance of an identified speaker. The fundamentals of the most commonly applied methods like long term statistics, vector quantization (VQ) and nearest neighbour method (NN) are included. An implementation of a vector quantization ina constructed speaker recognition system in a Matlab environment is shown and obtained results are discussed. The influence of several different speech parameters extracted from speaker utterances on identification accuracy is also included.Identification was done on Polish speech corpora ROBOT.