The use of hidden Markov models for speech recognition has become predominant in the last several years, as evidenced by the number of published papers and talks at major speech conferences. The reasons this method has become so popular are the inherent statistical (mathematically precise) framework; the ease and availability of training algorithms for cstimating the parameters of the models from finite training sets of speech data; the flexibility of the resulting recognition system in which one can easily change the size, type, or architecture of the models to suit particular words, sounds, and so forth; and the ease of implementation of the overall recognition system. In this expository article, we address the role of statistical methods in this powerful technology as applied to speech recognition and discuss a range of theoretical and practical issues that are as yet unsolved in terms of their importance and their effect on performance for different system implementations.

/pdf/hidden-markov-models-for-speech-recognition-1vqk5sdtvy.pdf

Hidden Markov models for speech recognition

A tutorial on signal processing in state-of-the-art speech recognition systems is presented, reviewing those techniques most commonly used. The four basic operations of signal modeling, i.e. spectral shaping, spectral analysis, parametric transformation, and statistical modeling, are discussed. Three important trends that have developed in the last five years in speech recognition are examined. First, heterogeneous parameter sets that mix absolute spectral information with dynamic, or time-derivative, spectral information, have become common. Second, similarity transform techniques, often used to normalize and decorrelate parameters in some computationally inexpensive way, have become popular. Third, the signal parameter estimation problem has merged with the speech recognition process so that more sophisticated statistical models of the signal's spectrum can be estimated in a closed-loop manner. The signal processing components of these algorithms are reviewed. >

/pdf/signal-modeling-techniques-in-speech-recognition-2s2amcdjv6.pdf

Signal modeling techniques in speech recognition

A new training procedure called multi-style training has been developed to improve performance when a recognizer is used under stress or in high noise but cannot be trained in these conditions Instead of speaking normally during training, talkers use different, easily produced, talking styles This technique was tested using a speech data base that included stress speech produced during a workload task and when intense noise was presented through earphones A continuous-distribution talker-dependent Hidden Markov Model (HMM) recognizer was trained both normally (5 normally spoken tokens) and with multi-style training (one token each from normal, fast, clear, loud, and question-pitch talking styles) The average error rate under stress and normal conditions fell by more than a factor of two with multi-style training and the average error rate under conditions sampled during training fell by a factor of four

Multi-style training for robust isolated-word speech recognition

The use of hidden Markov models (HMMs) in continuous speech recognition is reviewed. Markov models are presented as a generalization of their predecessor technology, dynamic programming. A unified view is offered in which both linguistic decoding and acoustic matching are integrated into a single, optimal network search framework. Advances in recognition architectures are discussed. The fundamentals of Viterbi beam search, the dominant search algorithm used today in speed recognition, are presented. Approaches to estimating the probabilities associated with an HMM model are examined. The HMM-supervised training paradigm is examined. Several examples of successful HMM-based speech recognition systems are reviewed. >

/pdf/continuous-speech-recognition-using-hidden-markov-models-2yiawm6wtz.pdf

Continuous speech recognition using hidden Markov models

The authors introduce a novel approach to modeling variable-duration phonemes, called the stochastic segment model. A phoneme X is observed as a variable-length sequence of frames, where each frame is represented by a parameter vector and the length of the sequence is random. The stochastic segment model consists of (1) a time warping of the variable-length segment X into a fixed-length segment Y called a resampled segment and (2) a joint density function of the parameters of X which in this study is a Gaussian density. The segment model represents spectra/temporal structure over the entire phoneme. The model also allows the incorporation in Y of acoustic-phonetic features derived from X, in addition to the usual spectral features that have been used in hidden Markov modeling and dynamic time warping approaches to speech recognition. The authors describe the stochastic segment model, the recognition algorithm, and an iterative training algorithm for estimating segment models from continuous speech. They present several results using segment models in two speaker-dependent recognition tasks and compare the performance of the stochastic segment model to the performance of the hidden Markov models. >

A stochastic segment model for phoneme-based continuous speech recognition

The performance of current speaker independent speech recognition technology is limited by the inadequacy of the measures of the speech data to discriminate between different speech sounds. In particular, two critical assumptions that underlie and limit most current recognition techniques are that: 1) speech data from different frames are statistically independent (e.g., there are no between-frame interactions); and 2) speech data statistics are independent of phonetic events (e.g., distance measures are fixed and independent of input or reference speech). In the context of speaker independent isolated digit recognition, improved recognition performance is demonstrated by: 1) explicitly modeling the correlation between spectral measurements of adjacent frames; and 2) using a distance measure which is a function of the recognition reference frame being used. A statistical model was created from a 2464 token database (2 tokens of each of 11 words "zero" through "nine" and "oh") for 112 speakers. Primary features include energy and filter bank amplitudes. Interspeaker variability was estimated by time aligning all training tokens and creating an ensemble of 224 feature vectors for each reference frame. Normal distributions were then estimated individually for each frame jointly with its neighbors. Testing was performed on a multidialect database of 2486 spoken digit tokens collected from 113 (different) speakers using maximum-likelihood decision methods. The substitution rate dropped from 1.7 to 1.4 percent with incorporation of between-frame statistics, and further to 0.6 percent with incorporation of frame-specific statistics in the likelihood model.

E. Bocchieri

Papers

Frame-specific statistical features for speaker independent speech recognition