Mean and variance adaptation within the MLLR framework

By Mark John Francis Gales, Philip C. Woodland. 1996

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

Parallel training of Deep Neural Networks with Natural Gradient and Parameter Averaging (2014, methodology): “…e.g. as reported in [15], include cepstral mean normalization (CMN), in which we subtract the mean of the input features; and fMLLR adaptation, also known as constrained MLLR adaptation [30], in which we use a baseline GMM system to compute a likelihood-maximizing linear transform of the features.”
Letter-Based Speech Recognition with Gated ConvNets (2017, background): “Most state of the art systems also perform speaker adaptation; iVectors compute a speaker embedding capturing both speaker and environment information (Xue et al., 2014), while fMMLR is a two-pass decoder technique which computes a speaker transform in the ﬁrst pass (Gales and Woodland, 1996).”
Acoustic Modelling from the Signal Domain Using CNNs (2016, methodology): “With traditional features, speaker adaptation is usually done using a generative framework that involves transforming features to a different space using fMLLR [14] etc.”
Uncertainty propagation through deep neural networks (2015, methodology): “The MFCC features have been post-processed using linear discriminant analysis (LDA) [18, 19], maximum likelihood linear regression (MLLR) [20, 21], and speaker adaptive training (SAT) [22].”
Context-Dependent Pre-Trained Deep Neural Networks for Large-Vocabulary Speech Recognition (2012, methodology): “Inspiration for such algorithms may come from the ANN-HMM literature (e.g., [72] and [73]) or the many successful adaptation techniques developed in the past decades for GMM-HMMs (e.g., MLLR [74], MAP [75], joint compensation of distortions [76], variable parameter HMMs [77]).”
Spoken language understanding using the Hidden Vector State Model (2006, methodology): “During decoding, incremental speaker adaptation based on maximum likelihood linear regression (MLLR) (Gales and Woodland, 1996) is performed with updating being performed every five input utterances.”
Transcribing Meetings With the AMIDA Systems (2012, methodology): “The systems outlined in this paper make use of adaptation in the form of cepstral mean and variance normalization per speaker, VTLN, andmaximum-likelihood linear regression (MLLR) [33].”
An Overview of Noise-Robust Automatic Speech Recognition (2014, background): “(45), it is better to also transform covariance parameters [120], [127] as (47) A two-stage optimization is usually used.”
Hermitian Polynomial for Speaker Adaptation of Connectionist Speech Recognition Systems (2013, methodology): “The best known example of transformation-based adaptation is the maximum likelihood linear regression (MLLR) framework [16], [17], in which an affine transformation is used to transform the mean and variance vectors of the Gaussian mixture densities in the original set of HMMs.”
Model-Based Approaches to Handling Uncertainty (2011, background): “As the KL divergence looks at complete distributions (rather than just the first-order moments in MMSE) changes in the uncertainty can be modelled with PCMLLR.”
Rapid speaker adaptation in eigenvoice space (2000, methodology): “[12], [31], [32] require significant amounts of adaptation data from the new speaker in order to perform better than a similar SI system.”
Improving deep neural network acoustic models using generalized maxout networks (2014, methodology): “Both are trained on top of features adapted with Constrained MLLR/fMLLR [27].”
Semi-tied covariance matrices for hidden Markov models (1999, methodology): “has a similar form to the variance transform described in [ 10 ].”
Machine Learning Paradigms for Speech Recognition: An Overview (2013, methodology): “adaptation techniques, such as MLLR [192], [193], SPLICE”