Citation Context

...ility space was introduced [1]. This total variability approach, avoids joint estimation of separate speaker and session spaces and factors, and is less reliant on application of score normalizations =-=[3, 1]-=-. In this paper we introduce a modification to the cosine similarity that does not require explicit score normalization, relying instead on simple mean and covariance statistics from a collection of i...

Citation Context

...rmalized following a standard normal distribution. 2.2. Total variability modeling The total variability space or i-vector approach concept was first introduced in the context of speaker verification =-=[9, 10]-=-. This approach was motivated by the success of the Joint Factor Analysis [11], which models both speaker and intersession subspaces separately. Unlike JFA, the total variability approach models all t...

Citation Context

... variability (parameters for a language can change from utterance to utterance because of the differences in channel, speaker, etc.). We will refer to the latter variability simply as channel variability. When the likelihood of a test utterance is evaluated for a certain language, the corresponding model is adapted to the channel of that test utterance. This is done by finding the point MAP (or ML) estimate of a low-dimensional latent variable vector - channel factors, which are coordinates in a highly channel-variable subspace of the model parameter space. Recently, systems based on iVectors [4, 16] have provided superior performance in speaker recognition. iVector is a fixedlength low-dimensional vector, which is extracted for each utterance based on the JFA-like idea of estimating latent variables corresponding to high variability subspace. The principal difference from JFA is that we are not interested in evaluating the adapted model. Instead, the latent variables - iVectors - are used as features for another (possibly very simple) classifier. Also, the underlying model for iVector extraction does not attempt to separate inter-class and channel variability. Instead, it considerers onl...

Citation Context

...ector extractor training—can be simplified and lead to a faster and memory more efficient code. The first assumption is that the GMM component alignment is constant across utterances and is given by the UBM GMM weights. The second assumption is that the i-vector extractor matrix can be linearly transformed so that its per-Gaussian components are orthogonal. We use PCA and HLDA to estimate this transform. Index Terms— speaker recognition, i-vectors, Joint Factor Analysis, PCA, HLDA 1. INTRODUCTION The i-vector systems have become the state-of-the-art technique in the speaker verification field [1]. They provide an elegant way of reducing the large-dimensional input data to a small-dimensional feature vector while retaining most of the relevant information. The technique was originally inspired by Joint Factor Analysis framework introduced in [2, 3]. The computational requirements for training the i-vector systems and estimating the i-vectors, however, are too high for certain types of applications. In this paper we propose simplifications to the original i-vector extraction and training schemes, which would dramatically decrease their complexity while retaining the recognition performa...

Citation Context

...trained using sufficient amount of data, a variety of phonemes are captured from the speaker enabling the model to better represent the speaker’s acoustic space while enhancing its discriminating ability. This has been especially true since Gaussian Mixture Models (GMM) were introduced for modeling the acoustic space for speaker recognition [1]. The subsequent research endeavors focused more on modeling the speaker dependent GMM mean super-vectors using various factor analysis methods, leading to techniques such as Eigenvoice [2, 3], Eigenchannel, Joint Factor Analysis (JFA) [4], and i-vector [5] based speaker recognition systems. Most of these systems were specifically designed to handle channel mismatch. The state-of-the-art systems generally utilize the i-vectors as utterance dependent features and use a Probabilistic Linear Discriminant Analysis (PLDA) [6, 7] model for classification, †This work is partly funded by European Community’s Seventh Framework Program (FP7/2007-2013) under grant agreement no. 238803. ?This project was funded by AFRL under contract FA8750-12-1-0188 and partially by the University of Texas at Dallas from the Distinguished University Chair in Telecommunicat...

Citation Context

...er recognition, linear discriminant analysis, i-vector, total variability, source variability 1. INTRODUCTION A speaker recognition framework based on i-vectors was recently developed by Dehak et al. =-=[1, 2]-=- offering superior recognition to the widely adopted joint factor analysis (JFA) approach [3]. This framework involves extracting i-vectors from a low-dimensional total variability subspace, enhancing...

Citation Context

...to perform well under noisy condition rather than clean (“not altered”) condition. 3. Recognition systems The systems developed in the I4U coalition were based on state-of-the-art: 1) i-vector system =-=[11]-=- with probabilistic linear discriminant analysis (PLDA) [12] modeling, or 2) Gaussian supervector representation and joint factor analysis (JFA) [13, 14], or support vector machine (SVM) modeling. All...

Citation Context

...LDA) is a generative model which is similar to the JFA approach. Rather than using GMM supervectors as the basis for factor modeling, PLDA models i-vector distributions for speaker verification [24], =-=[26]-=-, and model session and speaker variability within the i-vector space. The i-vector is a low-dimensional set of factors, and it is used to represent speaker and channel variability via factor loadings...

Citation Context

...[11], uncertainty in features is modeled by the variances in the Gaussian mixture models (GMMs) [12] and, recently, by subspace models of speaker and session variabilities in a supervector space [13]–=-=[19]-=-. However, if the MFCCs themselves are estimated with smaller variance, one should expect the subsequent speaker and session variability models to exhibit less random variation as well. Using lowvaria...

Citation Context

...g the different segments of speech belonging to the same speaker (clustering). [1] We develop an approach to diarization based on the successes of factor analysis-based methods in speaker recognition =-=[2]-=-, as well as diarization [3], [4]. Inspired by the ability of the Total Variability subspace to extract speaker-specific features on short segments of speech [2], [5], we propose a method for performi...