Semi-continuous Hidden Markov Models Research Articles

Signal bias removal by maximum likelihood estimation for robust telephone speech recognition

An acoustical mismatch between the training and the testing conditions of hidden Markov model (HMM)-based speech recognition systems often causes a severte degradation in the recognition performance. In telephone speech recognition, for example, undesirable signal components due to ambient noise and channel distortion, as well as due to different variations of telephone handsets render the recognizer unusable for real- world applications. This paper presents a signal bias removal (SBR) method based on maximum likelihood1 estimation for the minimization of these undesirable effects. The proposed method is readily applicable in various architectures, i.e., dis- crete (vector-quantization based), semicontinuous and continuous density HMM. In this paper, the SBR method, integrated into a discrete density HMM, is applied to telephone speech recognition where the contamination due to extraneous signal components is assumed to be unknown. To enable real-time implementation, a sequential method for the estimation of the bias is presented. Experimental results for speaker-independent connected digit recognition show a reduction in the per digit error rate by up to 41% and 14% during mismatched and matclhed training and testing conditions, respectively.

Discriminating observation probability (DOP) HMM for speaker verification

This paper describes the use of a multiple codebook semi-continuous hidden Markov model (SCHMM) automatic speaker verification (ASV) system, which uses a novel technique for discriminative hidden Markov modelling known as discriminative observation probabilities (DOP). DOP is not a discriminative training technique, it is a method of constructing what is effectively a discriminating model by contrasting two standard HMMs so as to improve discrimination between the classes that those models represent. This paper experimentally evaluates the use of DOP HMMs for ASV. The experimental evaluation is based on a text-dependent task using isolated digits. The database contains 24 true (client) speakers and 100 casual impostors, recorded over the public telephone network in the United Kingdom. For speaker verification, DOP has similarities to the so-called speaker normalisation technique but it is a more flexible approach and experimental results are presented here which show it to be superior, even in its simplest form. An argument is also made for preferring DOP models to discriminatively trained models although this is not evaluated experimentally. The DOP technique can easily be added to an existing HMM system and requires no additional training. The DOP concept can potentially be applied to other applications of HMMs to two-class classification problems. An equal error rate (EER) of 0.21% (using speaker specific thresholds, on a sequence of 12 isolated digits) was obtained using multiple codebook DOP models. This represents a reduction in EER of 38% compared to a common speaker normalisation technique.

Bayesian adaptive learning of the parameters of hidden Markov model for speech recognition

A theoretical framework for Bayesian adaptive training of the parameters of a discrete hidden Markov model (DHMM) and of a semi-continuous HMM (SCHMM) with Gaussian mixture state observation densities is presented. In addition to formulating the forward-backward MAP (maximum a posteriori) and the segmental MAP algorithms for estimating the above HMM parameters, a computationally efficient segmental quasi-Bayes algorithm for estimating the state-specific mixture coefficients in SCHMM is developed. For estimating the parameters of the prior densities, a new empirical Bayes method based on the moment estimates is also proposed. The MAP algorithms and the prior parameter specification are directly applicable to training speaker adaptive HMMs. Practical issues related to the use of the proposed techniques for HMM-based speaker adaptation are studied. The proposed MAP algorithms are shown to be effective especially in the cases in which the training or adaptation data are limited.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Multiple VQ hidden Markov modelling for speech recognition

In this paper a new variant of HMM, named Multiple VQ HMM (MVQHMM), is presented. Its main characteristic is the use of a separate codebook for each model. Procedures for training and probability evaluation of these models are described. The evaluation procedure combines the quantization distortions of the vector sequences with the discrete HMM generation probabilities. Comparative results on an isolated word recognition system are shown, between MVQHMM and discrete and semi-continuous HMM. These results show that using separate codebooks and including the quantization distortion in the decision criterion improve the performance of the system. Furthermore, the multiple VQ hidden Markov models seem to be more robust than the discrete and semi-continuous ones in relation to the inter-speaker variability of the recognition system.

HMM speaker verification with sparse training data on telephone quality speech

Speaker verification experiments using discrete and semi-continuous HMMs with telephone quality isolated digits are reported. The models were trained with varying numbers of tokens, giving equal error rates of 14% and 12%, respectively, on single isolated digits, and 4% and 2% on a sequence of 12 isolated digits.

On speaker-independent, speaker-dependent, and speaker-adaptive speech recognition

The DARPA Resource Management task is used as a domain for investigating the performance of speaker-independent, speaker-dependent, and speaker-adaptive speech recognition. The error rate of the speaker-independent recognition system, SPHINX, was reduced substantially by incorporating between-word triphone models additional dynamic features, and sex-dependent, semicontinuous hidden Markov models. The error rate for speaker-independent recognition was 4.3%. On speaker-dependent data, the error rate was further reduced to 2.6-1.4% with 600-2400 training sentences for each speaker. Using speaker-independent models, the authors studied speaker-adaptive recognition. Both codebooks and output distributions were considered for adaptation. It was found that speaker-adaptive systems outperform both speaker-independent and speaker-dependent systems, suggesting that the most effective system is one that begins with speaker-independent training and continues to adapt to users.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Phoneme classification using semicontinuous hidden Markov models

Speaker-dependent phoneme recognition experiments were conducted using variants of the semicontinuous hidden Markov model (SCHMM) with explicit state duration modeling. Results clearly demonstrated that the SCHMM with state duration offers significantly improved phoneme classification accuracy compared to both the discrete HMM and the continuous HMM; the error rate was reduced by more than 30% and 20%, respectively. The use of a limited number of mixture densities significantly reduced the amount of computation. Explicit state duration modeling further reduced the error rate.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Parameter re-estimation in semicontinuous hidden Markov modelling of speech with feedback to vector quantisation codebook

The parameters of the semicontinuous hidden Markov model (SCHMM) can be re-estimated by allowing the codebook to be updated, thus achieving an optimised codebook/model combination. With the optimised codebook, the SCHMM can offer improved recognition accuracy in comparison to both the continuous and the discrete hidden Markov model.

Performance comparison between semicontinuous and discrete hidden Markov models of speech

The semicontinuous hidden Markov model (SCHMM) of speech has been shown to offer improved performance in comparison to conventional discrete HMM. Experimental results with different vector quantisation levels, different probability density functions, and different numbers of tokens in the training set for the SCHMM are reported here to compare with conventional discrete HMM.