This paper deals with several cohort methods for score normalization in speaker verification syst... more This paper deals with several cohort methods for score normalization in speaker verification systems. At first, the reasons for score normalization are provided. Next, the principle of score normalization techniques based on Bayesian theorem are explained. The world, cohort, and unconstraint cohort normalization techniques are presented. A new normalization technique, unconstraint cohort extrapolated normalization, is introduced. Experiments on NIST 2002 corpus were performed in order to find which of the normalization methods give the best result. All experiments show that on-line cohort methods (especially unconstraint cohort extrapolated normalization) have outperformed the others. Finally, there is a discussion about time consumption of the on-line methods. An improvement for acceleration of these methods are proposed. The results of experiments on NIST 2002 data set showed same significant improvements of cohort and proposed cohort extrapolated normalization in comparison with the standard world normalization. The acceleration (4-times) on-line cohort methods is shown in another experiment.
The paper deals with training of speaker-clustered acoustic models. Various training techniques-M... more The paper deals with training of speaker-clustered acoustic models. Various training techniques-Maximum Likelihood, Discriminative Training and two adaptation based on the MAP and Discriminative MAP were tested in order to minimize an impact of speaker changes to the correct function of the recognizer when a response of the automatic cluster detector is delayed or incorrect. Such situation is very frequent e.g. in online subtitling of TV discussions (Parliament meetings). In our experiments the best cluster-dependent training procedure was discriminative adaptation which provided the best trade-off between recognition results with correct and non-correct cluster detector information.
It is well known that gender-dependent (male/female) acoustic models are more acoustically homoge... more It is well known that gender-dependent (male/female) acoustic models are more acoustically homogeneous and therefore give better recognition performance than single gender-independent model in the case where the gender is successfully detected or a priory known. Speakers do not need to be split to two groups only. An algorithm to make higher number of speaker clusters is described in this paper. Further, the paper deals with a problem how to use these gender-based or speaker-clustered acoustic models in a real-time LVCSR where information from an automatic cluster detector is often delayed or incorrect. Moreover, various ways, how to incorporate discriminative training methods into training of the speaker-clustered acoustic models, are discussed in the paper.
In any physico-chemical system, the number of degrees of freedom is defined primarily by number o... more In any physico-chemical system, the number of degrees of freedom is defined primarily by number of independent components and is typically much smaller than number of chemical individui (Atkins, de Paula 2008). Number of degree of freedom is number of coordinates in the state space. In dynamic system, such as living cell, the state space is charted into attraction zones of asymptotically stable attractors, namely limit cycles or strange attractors (Chaosbook). Attributes of the system are parameters of the system equation but proper description may be reduced to probability density function describing transition between attractor zones. First, attractor zones must be found from their characteristic time behavior. Number of degrees of freedom, as well as the characteristics of asymptotically stable objects (i.e. shapes, frequencies etc.) may only be determined experimentally. For that we need to identify self-assembling objects (i.e. phases, patterns etc.) and quantify them. In micro...
Entropy-based enhancement may be utilized also for analysis of monochromatic super resolved image... more Entropy-based enhancement may be utilized also for analysis of monochromatic super resolved images. Here, all features of the method appear unobscured.)Entropy-based enhancement emphasizes structural features of the image distinct from the bulk property. It enables logical sectioning of image features. The logic of the sectioning is different than that based on intensity image. The appearance of peaks in the entropy spectrum is not expected by theoretical models. It probably arises from extremely non-normal distribution of intensity probabilities.
Model λ = {ωm, µm, Σm} M m=1 p(ot |λ) = M m=1 ωmN (ot |µm, Σm) Multidimensional GMMs (dim(ot) = D... more Model λ = {ωm, µm, Σm} M m=1 p(ot |λ) = M m=1 ωmN (ot |µm, Σm) Multidimensional GMMs (dim(ot) = D) Diagonal vs. full covariance matrices Training-iterative EM algorithm E-step E-step: γm(ot) = ωmN (ot |µm,Σm) p(ot |λ) Outline GMMs CUDA Experiments Future work GPU architecture & CUDA GPU much more efficient on data-parallel tasks Fully parallel system operating with more than 100 of threads at once Algorithmic performance 10-50 times higher than 1 CPU core Memory bandwidth also significantly higher For maximum performance-several types of memories Single high-end GPU can fully replace CPU-only cluster(s) strongly (!) dependent on the implementation Outline GMMs CUDA Experiments Future work NVIDIA's Compute Unified Device Architecture (CUDA)
The work was supported by the Ministry of Education of the Czech Republic, project no. MŠMT 2C060... more The work was supported by the Ministry of Education of the Czech Republic, project no. MŠMT 2C06020. The paper deals with training of speaker-clustered acoustic models. Various training techniques-Maximum Likelihood, Discriminative Training and two adaptation based on the MAP and Discriminative MAP were tested in order to minimize an impact of speaker changes to the correct function of the recognizer when a response of the automatic cluster detector is delayed or incorrect. Such situation is very frequent e.g. in on-line subtitling of TV discussions (Parliament meetings).
Traditional discriminative training methods modify Hidden Markov Model (HMM) parameters obtained ... more Traditional discriminative training methods modify Hidden Markov Model (HMM) parameters obtained via a Maximum Likelihood (ML) criterion based estimator. In this paper, anti-models are introduced instead. The anti-models are used in tandem with ML models to incorporate a discriminative information from training data set and modify the HMM output likelihood in a discriminative way. Traditional discriminative training methods are prone to over-fitting and require an extra stabilization. Also, convergence is not ensured and usually "a proper" number of iterations is done. In the proposed anti-models concept, two parts, positive model and anti-model, are trained via ML criterion. Therefore, the convergence and the stability are ensured.
Annual Conference of the International Speech Communication Association, 2008
The article introduces an expert system for the speaker ver- ification task. Our main purpose was... more The article introduces an expert system for the speaker ver- ification task. Our main purpose was to design a tool for the combination of various speaker verification systems proposed for various operating conditions. First of all, the essential ideas are explained that made us design the expert system. Next sec- tion describes the structure of a rule-based expert system and
This paper deals with several cohort methods for score normalization in speaker verification syst... more This paper deals with several cohort methods for score normalization in speaker verification systems. At first, the reasons for score normalization are provided. Next, the principle of score normalization techniques based on Bayesian theorem are explained. The world, cohort, and unconstraint cohort normalization techniques are presented. A new normalization technique, unconstraint cohort extrapolated normalization, is introduced. Experiments on NIST 2002 corpus were performed in order to find which of the normalization methods give the best result. All experiments show that on-line cohort methods (especially unconstraint cohort extrapolated normalization) have outperformed the others. Finally, there is a discussion about time consumption of the on-line methods. An improvement for acceleration of these methods are proposed. The results of experiments on NIST 2002 data set showed same significant improvements of cohort and proposed cohort extrapolated normalization in comparison with the standard world normalization. The acceleration (4-times) on-line cohort methods is shown in another experiment.
The paper deals with training of speaker-clustered acoustic models. Various training techniques-M... more The paper deals with training of speaker-clustered acoustic models. Various training techniques-Maximum Likelihood, Discriminative Training and two adaptation based on the MAP and Discriminative MAP were tested in order to minimize an impact of speaker changes to the correct function of the recognizer when a response of the automatic cluster detector is delayed or incorrect. Such situation is very frequent e.g. in online subtitling of TV discussions (Parliament meetings). In our experiments the best cluster-dependent training procedure was discriminative adaptation which provided the best trade-off between recognition results with correct and non-correct cluster detector information.
It is well known that gender-dependent (male/female) acoustic models are more acoustically homoge... more It is well known that gender-dependent (male/female) acoustic models are more acoustically homogeneous and therefore give better recognition performance than single gender-independent model in the case where the gender is successfully detected or a priory known. Speakers do not need to be split to two groups only. An algorithm to make higher number of speaker clusters is described in this paper. Further, the paper deals with a problem how to use these gender-based or speaker-clustered acoustic models in a real-time LVCSR where information from an automatic cluster detector is often delayed or incorrect. Moreover, various ways, how to incorporate discriminative training methods into training of the speaker-clustered acoustic models, are discussed in the paper.
In any physico-chemical system, the number of degrees of freedom is defined primarily by number o... more In any physico-chemical system, the number of degrees of freedom is defined primarily by number of independent components and is typically much smaller than number of chemical individui (Atkins, de Paula 2008). Number of degree of freedom is number of coordinates in the state space. In dynamic system, such as living cell, the state space is charted into attraction zones of asymptotically stable attractors, namely limit cycles or strange attractors (Chaosbook). Attributes of the system are parameters of the system equation but proper description may be reduced to probability density function describing transition between attractor zones. First, attractor zones must be found from their characteristic time behavior. Number of degrees of freedom, as well as the characteristics of asymptotically stable objects (i.e. shapes, frequencies etc.) may only be determined experimentally. For that we need to identify self-assembling objects (i.e. phases, patterns etc.) and quantify them. In micro...
Entropy-based enhancement may be utilized also for analysis of monochromatic super resolved image... more Entropy-based enhancement may be utilized also for analysis of monochromatic super resolved images. Here, all features of the method appear unobscured.)Entropy-based enhancement emphasizes structural features of the image distinct from the bulk property. It enables logical sectioning of image features. The logic of the sectioning is different than that based on intensity image. The appearance of peaks in the entropy spectrum is not expected by theoretical models. It probably arises from extremely non-normal distribution of intensity probabilities.
Model λ = {ωm, µm, Σm} M m=1 p(ot |λ) = M m=1 ωmN (ot |µm, Σm) Multidimensional GMMs (dim(ot) = D... more Model λ = {ωm, µm, Σm} M m=1 p(ot |λ) = M m=1 ωmN (ot |µm, Σm) Multidimensional GMMs (dim(ot) = D) Diagonal vs. full covariance matrices Training-iterative EM algorithm E-step E-step: γm(ot) = ωmN (ot |µm,Σm) p(ot |λ) Outline GMMs CUDA Experiments Future work GPU architecture & CUDA GPU much more efficient on data-parallel tasks Fully parallel system operating with more than 100 of threads at once Algorithmic performance 10-50 times higher than 1 CPU core Memory bandwidth also significantly higher For maximum performance-several types of memories Single high-end GPU can fully replace CPU-only cluster(s) strongly (!) dependent on the implementation Outline GMMs CUDA Experiments Future work NVIDIA's Compute Unified Device Architecture (CUDA)
The work was supported by the Ministry of Education of the Czech Republic, project no. MŠMT 2C060... more The work was supported by the Ministry of Education of the Czech Republic, project no. MŠMT 2C06020. The paper deals with training of speaker-clustered acoustic models. Various training techniques-Maximum Likelihood, Discriminative Training and two adaptation based on the MAP and Discriminative MAP were tested in order to minimize an impact of speaker changes to the correct function of the recognizer when a response of the automatic cluster detector is delayed or incorrect. Such situation is very frequent e.g. in on-line subtitling of TV discussions (Parliament meetings).
Traditional discriminative training methods modify Hidden Markov Model (HMM) parameters obtained ... more Traditional discriminative training methods modify Hidden Markov Model (HMM) parameters obtained via a Maximum Likelihood (ML) criterion based estimator. In this paper, anti-models are introduced instead. The anti-models are used in tandem with ML models to incorporate a discriminative information from training data set and modify the HMM output likelihood in a discriminative way. Traditional discriminative training methods are prone to over-fitting and require an extra stabilization. Also, convergence is not ensured and usually "a proper" number of iterations is done. In the proposed anti-models concept, two parts, positive model and anti-model, are trained via ML criterion. Therefore, the convergence and the stability are ensured.
Annual Conference of the International Speech Communication Association, 2008
The article introduces an expert system for the speaker ver- ification task. Our main purpose was... more The article introduces an expert system for the speaker ver- ification task. Our main purpose was to design a tool for the combination of various speaker verification systems proposed for various operating conditions. First of all, the essential ideas are explained that made us design the expert system. Next sec- tion describes the structure of a rule-based expert system and
Uploads
Papers by Jan Vanek