Wyniki 1-1 spośród 1 dla zapytania: authorDesc:"Vjatcheslav V. KOVTUN"

Integration of hidden markov models in the automated speaker recognition system for critical use DOI:10.15199/48.2019.04.32

Czytaj za darmo! »

In speaker recognition systems (SRS) as a whole and in the automated speaker recognition system for critical use (ASRSCU) in particular, use the classical methods of pattern recognition theory, namely statistical simulation methods for describing the vectors of individual features of speech signals. Most often, in models of Gaussian mixtures [1, 2], artificial neural networks [3, 4] or support vector machines [5, 6, 7]. Less used are hidden Markov models [5, 6, 8, 9], which, however, together with Gaussian mixtures models, are very often used as part of speech recognition systems. Gaussian mixture models (GMM) are used in SRS to estimate the density of the probabilities of the variability of speech data due to moderately low computational cost of analysis and convergent adaptation algorithms [9, 10], in particular, the Expectancy-Maximization (EM) algorithm, the Maximum a Posteriori Probability (MAP) algorithm or Maximum Likelihood Linear Regression Maximization (MLLR) algorithm. However, the GMM has a low sensitivity to the variability of speech signal over time, which is usually compensated by detail for an adequate description of the individual features of speech, which leads to an increase in the sensitivity of the received features space to the presence in a phonogram of a speech signal a noises of the surrounding space. In their turn, the hidden Markov models (HMM) are statistical models that describe the analyzed system as a Markov process with unknown parameters [6] in order to determine the most probable state of the sequence of units of acoustic elements of speech signals based on pretrained models. For SRS, each state of the HMM is represented by different stable elements of speech (for example, the phonemes), and the time information is encoded by the permitted transitions between states. Thus, the speaker recognition using HMM is to determine for each speaker the optimal position between the sequence of the [...]

 Strona 1