https://www.cs.ubc.ca/~murphyk/Teaching/CS540-Spring10/projects/Yu-hsmm09.pdf

Hidden semi-Markov models

Voxceleb: Large-scale speaker verification in the wild

https://hal.inria.fr/hal-01399180/document

An analysis of environment, microphone and data simulation mismatches in robust speech recognition

Mine detection and classification using high-resolution sidescan sonar is a critical technology for mine counter measures (MCM). As opposed to the majority of techniques which require large training data sets, this paper presents unsupervised models for both the detection and the shadow extraction phases of an automated classification system. The detection phase is carried out using an unsupervised Markov random field (MRF) model where the required model parameters are estimated from the original image. Using a priori spatial information on the physical size and geometric signature of mines in sidescan sonar, a detection-orientated MRF model is developed which directly segments the image into regions of shadow, seabottom-reverberation, and object-highlight. After detection, features are extracted so that the object can be classified. A novel co-operating statistical snake (CSS) model is presented which extracts the highlight and shadow of the object. The CSS model again utilizes available a priori information on the spatial relationship between the highlight and shadow, allowing accurate segmentation of the object's shadow to be achieved.

/pdf/an-automatic-approach-to-the-detection-and-extraction-of-2fqnqbhw2h.pdf

An automatic approach to the detection and extraction of mine features in sidescan sonar

©2016 Association for Computational Linguistics. Moving from limited-domain natural language generation (NLG) to open domain is difficult because the number of semantic input combinations grows exponentially with the number of domains. Therefore, it is important to leverage existing resources and exploit similarities between domains to facilitate domain adaptation. In this paper, we propose a procedure to train multi-domain, Recurrent Neural Network-based (RNN) language generators via multiple adaptation steps. In this procedure, a model is first trained on counterfeited data synthesised from an out-of-domain dataset, and then fine tuned on a small set of in-domain utterances with a discriminative objective function. Corpus-based evaluation results show that the proposed procedure can achieve competitive performance in terms of BLEU score and slot error rate while significantly reducing the data needed to train generators in new, unseen domains. In subjective testing, human judges confirm that the procedure greatly improves generator performance when only a small amount of data is available in the domain.

/pdf/multi-domain-neural-network-language-generation-for-spoken-3rk6ja011o.pdf

Multi-domain Neural Network Language Generation for Spoken Dialogue Systems

This paper describes the Multi-Genre Broadcast (MGB) Challenge at ASRU 2015, an evaluation focused on speech recognition, speaker diarization, and "lightly supervised" alignment of BBC TV recordings. The challenge training data covered the whole range of seven weeks BBC TV output across four channels, resulting in about 1,600 hours of broadcast audio. In addition several hundred million words of BBC subtitle text was provided for language modelling. A novel aspect of the evaluation was the exploration of speech recognition and speaker diarization in a longitudinal setting — i.e. recognition of several episodes of the same show, and speaker diarization across these episodes, linking speakers. The longitudinal tasks also offered the opportunity for systems to make use of supplied metadata including show title, genre tag, and date/time of transmission. This paper describes the task data and evaluation process used in the MGB challenge, and summarises the results obtained.

/pdf/the-mgb-challenge-evaluating-multi-genre-broadcast-media-2rzs509bsz.pdf

The MGB challenge: Evaluating multi-genre broadcast media recognition

Copyright © 2015 ISCA. Recurrent neural network language models (RNNLMs) have recently become increasingly popular for many applications including speech recognition. In previous research RNNLMs have normally been trained on well-matched in-domain data. The adaptation of RNNLMs remains an open research area to be explored. In this paper, genre and topic based RNNLMadaptation techniques are investigated for a multi-genre broadcast transcription task. A number of techniques including Probabilistic Latent Semantic Analysis, Latent Dirichlet Allocation and Hierarchical Dirichlet Processes are used to extract show level topic information. These were then used as additional input to the RNNLM during training, which can facilitate unsupervised test time adaptation. Experiments using a state-of-theart LVCSR system trained on 1000 hours of speech and more than 1 billion words of text showed adaptation could yield perplexity reductions of 8% relatively over the baseline RNNLM and small but consistent word error rate reductions.

Recurrent neural network language model adaptation for multi-genre broadcast speech recognition

Hidden Markov fields (HMF) are widely used in image processing. In such models, the hidden random field of interest X=(Xs) is a Markov field, and the distribution p(y/x) of the observed random field Y=(Ys) conditional on X is given by the product of p(ys/xs), with s in the set of pixels. The posterior distribution p(x/y) is then a Markov distribution, which affords different Bayesian processing. However, when dealing with the segmentation of images containing numerous classes with different textures, the simple form of the distribution p(y/x) above is insufficient and has to be replaced by a Markov field distribution. This poses problems, because taking p(y/x) Markovian implies that the posterior distribution p(x/y), whose Markovianity is needed to use Bayesian techniques, may no longer be a Markov distribution, and so different model approximations must be made to remedy this. This drawback disappears when considering directly the Markovianity of (X, Y); in these recent 'Pairwise Markov Fields (PMF) models, both p(y/x) and p(x/y) are then Markovian, the first one allowing us to model textures, and the second one allowing us to use Bayesian restoration without model approximations. In this paper we generalize the PMF to Triplet Markov Fields (TMF) by adding a third random field U=(Us) and considering the Markovianity of (X, U, Y). We show that in TMF X is still estimable from Y by Bayesian methods. The parameter estimation with Iterative Conditional Estimation (ICE) is specified and we give some numerical results showing how the use of TMF can improve the classical HMF based segmentation.© (2003) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

/pdf/statistical-image-segmentation-using-triplet-markov-fields-2ylfan42ie.pdf

Statistical image segmentation using Triplet Markov fields

http://www-public.int-evry.fr/~pieczyn/A50.pdf

Unsupervised segmentation of randomly switching data hidden with non-Gaussian correlated noise

This paper addresses the problem of unsupervised Bayesian hidden Markov chain restoration. When the hidden chain is stationary, the classical "Hidden Markov Chain" (HMC) model is quite efficient, and associated unsupervised Bayesian restoration methods using the "Expectation-Maximization" (EM) algorithm work well. When the hidden chain is non stationary, on the other hand, the unsupervised restoration results using the HMC model can be poor, due to a bad match between the real and estimated models. The novelty of this paper is to offer a more appropriate model for hidden nonstationary Markov chains, via the theory of evidence. Using recent results relating to Triplet Markov Chains (TMCs), we show, via simulations, that the classical restoration results can be improved by the use of the theory of evidence and Dempster-Shafer fusion. The latter improvement is performed in an entirely unsupervised way using an original parameter estimation method. Some application examples to unsupervised image segmentation are also provided.

/pdf/unsupervised-restoration-of-hidden-nonstationary-markov-1tzvcrn0d9.pdf

Pierre Lanchantin

Papers

The MGB challenge: Evaluating multi-genre broadcast media recognition

Recurrent neural network language model adaptation for multi-genre broadcast speech recognition

Statistical image segmentation using Triplet Markov fields

Unsupervised segmentation of randomly switching data hidden with non-Gaussian correlated noise

Unsupervised restoration of hidden nonstationary Markov chains using evidential priors