Semi-tied covariance matrices for hidden Markov models

By Mark John Francis Gales. 1999

Subfields: Model Architectures, Probabilistic & Bayesian Methods

How researchers see this paper

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

In Model Architectures

In Probabilistic & Bayesian Methods

Citations in the corpus

Librispeech: An ASR corpus based on public domain audio books (2015, methodology): “For this first decoding pass we use a triphone model discriminatively trained with Boosted MMI [12], based on MFCC [13] features processed with frame-splicing over 7 frames, followed by LDA, followed by a global semi-tied covariance (STC) transform [14].”
Analysis of Speaker Adaptation Algorithms for HMM-Based Speech Synthesis and a Constrained SMAPLR Adaptation Algorithm (2009, background): “In [56], it is reported that full covariance modeling using semi-tied covariance [57] (also known as maximum-likelihood linear transformation [58]) has an effect on the parameter-generation algorithm considering global variance (GV) [51].”
Gaussianization (2000, methodology): “Although our iterative scheme may be slightly more expensive per iter ation than standard numerical optimization techniques such as Attias’ algorithm, in prac tice it converges after very few iterations, as observed in Gales (1999) [3].”
Model-Based Approaches to Handling Uncertainty (2011, methodology): “Due to this reason, and the added problem of delta and delta-delta parameters, featureenhancement schemes based on stereo training [3] are used to combine noise robustness with state-of-the-art front-end processing such as semi-tied transforms [19] and fMPE [51].”
Discriminative Input Stream Combination for Conditional Random Field Phone Recognition (2009, background): “covariance matrices [6], which look for structural dependencies between the feature dimensions.”
Investigation of unsupervised adaptation of DNN acoustic models with filter bank input (2014, unknown): “Since FBANK features are highly correlated with each other, it is essential to use full covariance matrices or semi-tied covariance (STC) transforms [17].”
Eigenvoice modeling with sparse training data (2005, methodology): “) More ambitiously, one could replace each Gaussian in a HMM output distribution by an independent components analyzer, tying the mixing matrices in the spirit of [22].”
Semi-supervised training of Deep Neural Networks (2013, methodology): “These features are then mean/variance normalized, spliced by +/- 4 frames next to the central frame and projected down to 40 dimensions using linear discriminant analysis (LDA) and single semi-tied covariance (STC) transform [20].”
Improvements to Deep Convolutional Neural Networks for LVCSR (2013, methodology): “Semi-tied covariance matrices (STCs) [14] have been used to decorrelate the feature space so that it can be modeled by diagonal Gaussians.”
Automatic speech recognition and speech variability: A review (2007, methodology): “Finally, LDA and HDA can be combined with maximum likelihood linear transform (MLLT) (Gopinath, 1998), which is identical to semi-tied covariance matrices (STC) (Gales, 1999).”
The subspace Gaussian mixture model - A structured model for speech recognition (2011, methodology): “We use a Constrained MLLR (CMLLR)3 feature transformation, as in Gales (1997), whereby we replace the feature x with the transformed feature x′ = A(s)x+ b(s) (8) where s is the speaker index.”
Speaker adaptation of context dependent deep neural networks (2013, methodology): “The speaker independent, real-time GMM system uses PLP features [20], semi-tied covariances [21] and linear discriminant analysis for dimensionality reduction of”
Accurate and compact large vocabulary speech recognition on mobile devices (2013, methodology): “Semi-tied covariances [6] are used to further diagonal-ize the LDA transformed features.”
The Kaldi Speech Recognition Toolkit (2011, background): “We support most commonly used feature extraction approaches: e.g. VTLN, cepstral mean and variance normalization, LDA, STC/MLLT, HLDA, and so on.”
Optimization Techniques to Improve Training Speed of Deep Neural Networks for Large Speech Tasks (2013, unknown): “semi-tied covariance (STC) transform [35].”
Sequence-discriminative training of deep neural networks (2013, methodology): “We also use a single semi-tied covariance (STC) transform [24] on the features obtained using LDA.”
Feature optimized DPGMM clustering for unsupervised subword modeling: A contribution to zerospeech 2017 (2017, background): “• Maximum likelihood linear transforms (MLLT) to reduce complexity and to de-correlate features [13, 14], • Feature-space maximum likelihood linear regression (fMLLR) to reduce speaker variability within speech features [15, 16],”
Nonlinear Compensation Using the Gauss–Newton Method for Noise-Robust Speech Recognition (2012, methodology): “Typical schemes include semi-tied covariance (STC) models [36], extended maximum-likelihood linear transform (EMLLT) [37], subspace for precision and mean (SPAM) models [38], and factor analyzed covariance models [39].”