State mixture modelling applied to speech recognition

2003

We present a generative factor analyzed hidden Markov model (GFA-HMM) for automatic speech recognition. In a traditional HMM, the observation vectors are represented by mixture of Gaussians (MoG) that are dependent on discrete-valued hidden state sequence. The GFA-HMM introduces a hierarchy of continuousvalued latent representation of observation vectors, where latent vectors in one level are acoustic-unit dependent and the latent vectors in a higher level are acoustic-unit independent. An expectation maximization (EM) algorithm is derived for maximum likelihood parameter estimation of the model. The GFA-HMM can achieve a much more compact representation of the intra-frame statistics of observation vectors than traditional HMM. We conducted an experiment to show that the GFA-HMM can achieve better performances over traditional HMM with the same amount of training data but much smaller number of model parameters.

Abstract—Speech Recognition is a process of transcribing speech to text. Phoneme based modeling is used where in each phoneme is represented by Continuous Density Hidden Markov Model. Mel Frequency Cepstral Coefficients (MFCC) are extracted from speech signal, delta and double-delta features representing the temporal rate of change of features are added which considerably improves the recognition accuracy. Each phoneme is represented by tristate Hidden Markov Model(HMM) with each state being represented by Continuous Density Gaussian model. As single mixture gaussian model do not represent the distribution of feature vectors in a better way, mixture splitting is performed successively in stages to eight mixture gaussian components. The multi-gaussian monophone models so generated do not capture all the variations of a phone with respect to its context, context dependent triphone models are build and the states are tied using decision tree based clustering. It is observed that recognition accuracy increases as the number of mixture components is increased and it works well for tied-state triphone based HMMs for large vocabulary. TIMIT Acoustic-Phonetic Continuous Speech Corpus is used for implementation. Recognition accuracy is also tested for our recorded speech.

18th International Conference on Pattern Recognition (ICPR'06), 2006

Models (HMMs), which represent the temporal dynamics of speech very efficiently, and Gaussian mixture models, which do non-optimally the classification of speech into single speech units (phonemes). In this paper we use parallel mixtures of Support Vector Machines (SVMs) for classification by integrating this method in a HMM-based speech recognition system. SVMs are very appealing due to their association with statistical learning theory and have already shown good results in pattern recognition and in continuous speech recognition. They suffer however from the effort for training which scales at least quadratic with respect to the number of training vectors. The SVM mixtures need only nearly linear training time making it easier to deal with the large amount of speech data. In our hybrid system we use the SVM mixtures as acoustic models in a HMM-based decoder. We train and test the hybrid system on the DARPA Resource Management (RM1) corpus, showing better performance than HMM-based decoder using Gaussian mixtures.

International Journal of Fuzzy Systems

1 This paper applies fuzzy vector quantization (FVQ) to the modeling of Discrete Hidden Markov Model (DHMM) and then to improve the speech recognition rate for the Mandarin speech. Vector quantization based on a codebook is a fundamental process to recognize the speech signal by DHMM. A codebook will be first trained by K-means algorithms using Mandarin training speech. Then, based on the trained codebook, the speech features are quantized by the fuzzy sets defined on each vectors of the code-book. Subsequently, the quantized speech features are statistically applied to train the model of DHMM for the speech recognition. All the speech features to be recognized should go through the FVQ based on the fuzzy codebook before being fed into the DHMM model for recognition. Experimental results in this paper shows that the speech recognition rate can be improved by using FVQ algorithm to train the model of DHMM.

2008 IEEE International Conference on Acoustics, Speech and Signal Processing, 2008

In this paper, a new acoustic model called Time-Inhomogeneous Hidden Bernoulli Model (TI-HBM) is introduced as an alternative to Hidden Markov Model (HMM) in automatic speech recognition. Contrary to HMM, the state transition process in TI-HBM is not a Markov process; rather it is an independent (generalized Bernoulli) process. This difference leads to elimination of dynamic programming at state-level in TI-HBM decoding process. Thus, the computational complexity of TI-HBM for Probability Evaluation and State Estimation is () NL (instead of 2 () N L in the HMM case). As a new framework for phone duration modeling, TI-HBM is able to model acoustic-unit duration (e.g. phone duration) by using a built-in parameter named survival probability. Similar to the HMM case, three essential problems in TI-HBM have been solved. An EM-algorithm based method has been proposed for training TI-HBM parameters. Experiments in phone recognition for Persian (Farsi) spoken language show that the TI-HBM has some advantages over HMM (e.g. more simplicity and increased speed in recognition phase), and also outperforms HMM in terms of phone recognition accuracy.

Interspeech 2012, 2012

IEICE Transactions on Information and Systems, 2006

In recent years, the number of studies investigating new directions in speech modeling that goes beyond the conventional HMM has increased considerably. One promising approach is to use Bayesian Networks (BN) as speech models. Full recognition systems based on Dynamic BN as well as acoustic models using BN have been proposed lately. Our group at ATR has been developing a hybrid HMM/BN model, which is an HMM where the state probability distribution is modeled by a BN, instead of commonly used mixtures of Gaussian functions. In this paper, we describe how to use the hybrid HMM/BN acoustic models, especially emphasizing some design and implementation issues. The most essential part of HMM/BN model building is the choice of the state BN topology. As it is manually chosen, there are some factors that should be considered in this process. They include, but are not limited to, the type of data, the task and the available additional information. When context-dependent models are used, the state-level structure can be obtained by traditional methods. The HMM/BN parameter learning is based on the Viterbi training paradigm and consists of two alternating steps-BN training and HMM transition updates. For recognition, in some cases, BN inference is computationally equivalent to a mixture of Gaussians, which allows HMM/BN model to be used in existing decoders without any modification. We present two examples of HMM/BN model applications in speech recognition systems. Evaluations under various conditions and for different tasks showed that the HMM/BN model gives consistently better performance than the conventional HMM.

2004

Most of the current state-of-the-art speech recognition systems are based on HMMs which usually use mixture of Gaussian functions as state probability distribution model. It is a common practice to use EM algorithm for Gaussian mixture parameter learning. In this case, the learning is done in a "blind", data-driven way without taking into account how the speech signal has been produced and which factors it depends on. In this paper, we describe the hybrid HMM/BN acoustic modeling framework, where, in contract to the conventional mixture of Gaussians, HMM state probability distribution is modeled by a Bayesian Network, hence the name is HMM/BN. Temporal speech characteristics are still governed by the HMM state transitions, but the state output likelihood is inferred from the BN. This allows for very flexible and consistent models of the state probability distributions which can easily integrate different speech parameterizations. BN can represent various speech features and environment conditions and their underlying dependencies. We show that the conventional HMM is a special case of HMM/BN model which we regard as a generalization of the HMM. The HMM/BN parameter learning is based on the Viterbi training paradigm and consists of two alternating steps -BN training and HMM transition probabilities update. For recognition, in some cases, BN inference is computationally equivalent to mixture of Gaussians which allows HMM/BN model to be used in existing HMM decoders. We present several examples of HMM/BN model application in speech recognition systems. Evaluations under various conditions and for different tasks showed that the HMM/BN model gives consistently better performance that the standard mixture of Gaussians HMM.

18th International Conference of the North American Fuzzy Information Processing Society - NAFIPS (Cat. No.99TH8397)

This paper proposes a Frzzy approach to the hidden Markov model (HMM) method called the futzr HMM for speech and speaker recognition The fuzzy HMM algorithm is regarded as an application of the fuzzy expectation-maximisation (EM) algorithm to the Baum-Welch algorithm in the HMM. Speech and speaker recognition experiments using the Texas Instruments p 4 6) speech data corpus show better results for fuuy HMMs compared with conventioml HMMs.

State of the art automatic speech recognition system uses Mel frequency cepstral coefficients as feature extractor along with Gaussian mixture model for acoustic modeling but there is no standard value to assign number of mixture component in speech recognition process.Current choice of mixture component is arbitrary with little justification. Also the standard set for European languages can not be used in Hindi speech recognition due to mismatch in database size of the languages.The parameter estimation with too many or few component may inappropriately estimate the mixture model. Therefore, number of mixture is important for initial estimation of expectation maximization process. In this research work, we estimate number of Gaussian mixture component for Hindi database based upon the size of vocabulary.Mel frequency cepstral feature and perceptual linear predictive feature along with its extended variations with delta-delta-delta feature have been used to evaluate this number based on optimal recognition score of the system . Comparitive analysis of recognition performance for both the feature extraction methods on medium size Hindi database is also presented in this paper.HLDA has been used as feature reduction technique and also its impact on the recognition score has been highlighted here.