Joint Factor Analysis of Speaker and Session Variability: Theory and Algorithms

By Patrick Kenny. 2006

452 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

VoxCeleb: A Large-Scale Speaker Identification Dataset (2017, methodology): “The state of the art in more recent times involves both the use of joint factor analysis (JFA) based methods which model speaker and channel subspaces separately [9], and i-vectors which attempt to model both subspaces into a single compact, low-dimensional space [10].”
Text dependent speaker verification using a small development set (2012, methodology): “Our Joint Factor Analysis (JFA)-based system is inspired by the theory described thoroughly in [2].”
Comparison of Speaker Recognition Approaches for Real Applications (2011, methodology): “The Joint Factor Analysis models have been trained following the guidelines of [2], [3] and [4] with some slight variations.”
Robust Speaker Identification in Noisy and Reverberant Conditions (2014, background): “State-of-the-art systems include joint factor analysis [13] and i-vector based techniques [3].”
Speaker adaptation of neural network acoustic models using i-vectors (2013, background): “Under the assumption of a fixed (soft) alignment of frames to mixture components, it can be shown that this posterior distribution is Gaussian [12] p(w|{xt(s)}) = N (w;L −1(s) K ∑ k=1 T T k Σ −1 k θk(s),L −1(s)) (4) with precision matrix L(s) of size M × M expressed as L(s) = I + K ∑ k=1 γk(s)T T k ”
Non-parallel voice conversion using i-vector PLDA: towards unifying speaker verification and transformation (2017, methodology): “Unlike the joint factor analysis (JFA) model [19], the i-vector model does not distinguish between speaker and other signal variations.”
Neural Network Bottleneck Features for Language Identification (2014, methodology): “The technique was originally inspired by Joint Factor Analysis framework introduced in [22].”
Speaker recognition by means of Deep Belief Networks (2013, methodology): “A Maximum-Likelihood estimate of matrix T is obtained by minor modifications of the Joint Factor Analysis approach [16][17].”
Deep bottleneck features for i-vector based text-independent speaker verification (2015, methodology): “The technique is inspired by joint factor analysis (JFA) framework [5, 12], where the Gaussian mixture model (GMM) mean is factorized into the speaker and channel factors using separate subspaces.”
An overview of text-independent speaker recognition: From features to supervectors (2010, methodology): “One possible way is to first estimate V followed by U and D [110, 112].”
Joint Factor Analysis Versus Eigenchannels in Speaker Recognition (2007, methodology): “There are several ways of constructing a likelihood function for using joint factor analysis models in speaker recognition [4].”
Bayesian Speaker Verification with Heavy-Tailed Priors (2010, background): “As explained in [4] and [13] (Section IIID), a direct calculation of the posterior of the hidden variables is still feasible in this case but it entails inverting an R × R block matrix.”
Spoofing and countermeasures for speaker verification: A survey (2015, methodology): “In particular, joint factor analysis (JFA) (Kenny, 2006) can improve ASV performance by incorporating distinct speaker and channel subspace models.”
New Developments in Voice Biometrics for User Authentication (2011, methodology): “Our JFA-based system is inspired by the theory described thoroughly in [1].”
Voxceleb: Large-scale speaker verification in the wild (2020, methodology): “The state of the art in more recent times involves both the use of joint factor analysis (JFA) based methods which model speaker and channel subspaces separately (Kenny, 2005), and i-vectors which attempt to model both subspaces into a single compact, low-dimensional space (Dehak et al., 2011).”
Full-covariance UBM and heavy-tailed PLDA in i-vector speaker verification (2011, methodology): “Model hyper-parameters T are estimated using the same EM algorithm as in case of JFA [6].”
What Does the Speaker Embedding Encode? (2017, unknown): “The pursuit of fixed-length vector representations gained momentum with the introduction of Joint Factor Analysis (JFA) [5], which modeled speaker and channel factors in separate sub-spaces within the super-vector space.”
Unsupervised Learning for Expressive Speech Synthesis (2017, methodology): “More details can be found in Kenny (2005). I-vectors derive from the factor analysis, however, in comparison to the JFA approach, the speaker and the channel variability are combined, since it has been reported by Dehak (2009); Dehak et al.”
Analysis of i-vector Length Normalization in Speaker Recognition Systems (2011, methodology): “In principle, an i-vector extractor is supposed to generate i-vectors that behave in this way (especially if a Minimum Divergence [7] step is used).”