1. Technical Field
This disclosure relates to the training of hidden Markov models.
2. Description of Related Art
A hidden Markov model (HMM) is a probabilistic model for time series data. Its many applications include speech recognition, computational biology, see A. Krogh, M. Brown, I. S. Mian, K. Sjolander, and D. Haussler, “Hidden Markov models in computational biology: Applications to protein modeling,” Journal of molecular biology, vol. 235, no. 5, pp. 1501-1531, 1994, S. R. Eddy, “Profile hidden Markov models.,” Bioinformatics, vol. 14, no. 9, pp. 755-763, 1998, K. Karplus, C. Barrett, and R. Hughey, “Hidden Markov models for detecting remote protein homologies.,” Bioinformatics, vol. 14, no. 10, pp. 846-856, 1998; computer vision, see J. Yamato, J. Ohya, and K. Ishii, “Recognizing human action in time-sequential images using hidden Markov model,” in Proc. CVPR. IEEE, 1992, pp. 379-385., M. Brand, N. Oliver, and A. Pentland, “Coupled hidden Markov models for complex action recognition,” in Proc. CVPR. IEEE, 1997, pp. 994-999, wavelet-based signal processing, M. S. Crouse, R. D. Nowak, and R. G. Baraniuk, “Wavelet-based statistical signal processing using hidden Markov models,” IEEE Transactions on Signal Processing, vol. 46, no. 4, pp. 886-902, 1998, control theory, R. J. Elliott, L. Aggoun, and J. B. Moore, Hidden Markov models: Estimation and Control, vol. 29, Springer, 1994; Bayesian reasoning, see C. M. Bishop, Pattern Recognition and Machine Learning, Springer, 2006, and spatial and time series processing.
HMMs are especially widespread in speech processing and recognition. Many popular speech recognition toolkits use HMMs: Hidden Markov Model Toolkit (HTK); see S. Young, G. Evermann, D. Kershaw, G. Moore, J. Odell, D. Ollason, V. Valtchev, and P. Woodland, “The HTK book,” Cambridge University Engineering Department, vol. 3, 2002; Sphinx, W. Walker, P. Lamere, P. Kwok, B. Raj, R. Singh, E. Gouvea, P. Wolf, and J. Woelfel, “Sphinx-4: A flexible open source framework for speech recognition,” 2004, SONIC B; Pellom and K. Hacioglu, “Recent improvements in the CU SONIC ASR system for noisy speech: The SPINE task,” in Proc. ICASSP. IEEE, 2003, vol. 1, pp. 1-4; RASR D. Rybach, C. Gollan, G. Heigold, B. Hoffmeister, J. Lf, R. Schlter, and H. Ney, “The RWTH Aachen University open source speech recognition system,” in Proc. Interspeech, 2009, pp. 2111-2114; Kaldi D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Mothcek, Y. Qian, P. Schwarz, et al., “The Kaldi speech recognition toolkit,” in Proc. ASRU, 2011; Attila H. Soltau, G. Saon, and B. Kingsbury, “The IBM Attila speech recognition toolkit,” in Proc. SLT. IEEE, 2010, pp. 97-102; Y. Chow, M. Dunham, O. Kimball, M. Krasner, G. Kubala, J. Makhoul, P. Price, S. Roucos, and R. Schwartz, “BYBLOS: The BBN continuous speech recognition system,” in Proc. ICASSP. IEEE, 1987, vol. 12, pp. 89-92]; and Watson V. Goffin, C. Allauzen, E. Bocchieri, D. Hakkani-Tur, A. Ljolje, S. Parthasarathy, M. Rahim, G. Riccardi, and M. Saraclar, “The AT&T Watson speech recognizer,” in Proc. ICASSP, 2005, pp. 1033-1036.
HMMs relate to neural networks in several ways. The forward algorithm of Baum-Welch HMM training resembles the training of some recurrent neural networks. J. S. Bridle, “Alpha-Nets: A recurrent neural network architecture with a hidden Markov model interpretation,” Speech Communication, vol. 9, no. 1, pp. 83-92, 1990. Modern automatic speech recognition also relies on both HMMs and neural networks. Neural-HMM hybrid architectures have improved the performance of speech recognition in many cases.
Earlier efforts, A. Krogh, M. Brown, I. S. Mian, K. Sjolander, and D. Haussler, “Hidden Markov models in computational biology: Applications to protein modeling,” Journal of molecular biology, vol. 235, no. 5, pp. 1501-1531, 1994, S. R. Eddy et al., “Multiple alignment using hidden Markov models,” in Proc. ISMB, 1995, vol. 3, pp. 114-120, used annealed noise to perturb the model parameters and to pick an alignment path between HMM states and the observed speech data.
Training hidden Markov models is computational intensive, and can take a great deal of time. The prior art has not addressed this drawbrack of such training in any fundamental, data-dependent way.
A learning computer system may estimate unknown parameters and states of a stochastic or uncertain system having a probability structure. The system may include a data processing system that may include a hardware processor. The system may: receive data; generate random, chaotic, fuzzy, or other numerical perturbations of the data, one or more of the states, or the probability structure; estimate observed and hidden states of the stochastic or uncertain system using the data, the generated perturbations, previous states of the stochastic or uncertain system, or estimated states of the stochastic or uncertain system; and cause perturbations or independent noise to be injected into the data, the states, or the stochastic or uncertain system so as to speed up training or learning of the probability structure and of the system parameters or the states.
The data processing system may cause the perturbations of the data, states, or probability structure to speed up training of a hidden Markov model.
The perturbations of the data, states, or probability structure may satisfy the Noisy Expectation Maximization (NEM) condition.
The system may make HMM state or parameter estimates and the perturbations may be used to improve the accuracy of the estimates.
The data processing system may cause the perturbations of the data to speed up training of the hidden Markov model.
The perturbations may train or update one or more mixture models in the probability structure.
One or more of the mixture models may include a Gaussian mixture model.
The injected perturbations or noise may have a rate that decays as the training proceeds.
The injection may include adding, multiplying, exponentiating the data, states, or probability structure with the perturbations or independent noise.
A non-transitory, tangible, computer-readable storage media may contain a program of instructions that cause a computer system comprising a data processing system that includes a hardware processor running the program of instructions to estimate unknown parameters and states of a stochastic or uncertain system having a probability structure using one or more of the approaches described herein.
These, as well as other components, steps, features, objects, benefits, and advantages, will now become clear from a review of the following detailed description of illustrative embodiments, the accompanying drawings, and the claims.
The drawings are of illustrative embodiments. They do not illustrate all embodiments. Other embodiments may be used in addition or instead. Details that may be apparent or unnecessary may be omitted to save space or for more effective illustration. Some embodiments may be practiced with additional components or steps and/or without all of the components or steps that are illustrated. When the same numeral appears in different drawings, it refers to the same or like components or steps.
Illustrative embodiments are now described. Other embodiments may be used in addition or instead. Details that may be apparent or unnecessary may be omitted to save space or for a more effective presentation. Some embodiments may be practiced with additional components or steps and/or without all of the components or steps that are described.
Noise can speed training in hidden Markov models (HMMs). A Noisy Expectation-Maximization (NEM) algorithm may inject noise when learning the maximum-likelihood estimate of the HMM parameters because an underlying Baum-Welch training algorithm is a special case of the Expectation-Maximization (EM) algorithm. The NEM theorem may give a sufficient condition for such an average noise boost. The condition may be a simple quadratic constraint on the noise when the HMM uses a Gaussian mixture model at each state. The noise benefit condition is derived for the case of an HMM with a Binomial mixture model (BMM) at each state. Simulations show that a noisy HMM converges faster than a noiseless HMM on the TIMIT data set.
Careful noise injection can speed the training process for a hidden Markov model (HMM). The proper noise appears to help the training process explore less probable regions of the parameter space. The new system is called a noisy HMM or NHMM.
The NHMM may be a special case of a recent noisy EM (NEM) model, see O. Osoba, S. Mitaim, and B. Kosko, “Noise Benefits in the Expectation-Maximization Algorithm: NEM theorems and Models,” in The International Joint Conference on Neural Networks (IJCNN). IEEE, 2011, pp. 3178-3183; and 0. Osoba, S. Mitaim, and B. Kosko, “The Noisy Expectation-Maximization Algorithm,” Fluctuation and Noise Letters, vol. 12, no. 13, 1350012-1 1350012-30, September 2013. The underlying NEM theorem states that the noise-enhanced EM algorithm converges faster on average to the maximum-likelihood optimum than does the noiseless EM algorithm if the noise obeys a positivity condition. The condition reduces to a quadratic constraint on the injected noise in the special but important case of a Gaussian mixture model. The NEM algorithm gives rise to the NHMM because the Baum-Welch algorithm that trains the HMM parameters is itself a special case of the EM algorithm. See L. E. Baum, T. Petrie, G. Soules, and N. Weiss, “A maximization technique occurring in the statistical analysis of probabilistic functions of Markov chains,” The Annals of Mathematical Statistics, pp. 164-171, 1970. Theorem 1 below states the corresponding sufficient condition for an HMM noise boost. This is a type of “stochastic resonance” effect where a small amount of noise improves the performance of a nonlinear system while too much noise harms the system.
The simulations below confirm the theoretical prediction that proper injection of noise can improve speech recognition. This is a deliberate use of noise injection in the speech data itself. Earlier efforts, A. Krogh, M. Brown, I. S. Mian, K. Sjolander, and D. Haussler, “Hidden Markov models in computational biology: Applications to protein modeling,” Journal of molecular biology, vol. 235, no. 5, pp. 1501-1531, 1994, and S. R. Eddy et al., “Multiple alignment using hidden Markov models,” in Proc. ISMB, 1995, vol. 3, pp. 114-120, used annealed noise to perturb the model parameters and to pick an alignment path between HMM states and the observed speech data. These earlier efforts neither added noise to the speech data nor found any theoretical guarantee of a noise benefit.
The Noisy Expectation-Maximization (NEM) algorithm (see O. Osoba, S. Mitaim, and B. Kosko, “Noise Benefits in the Expectation-Maximization Algorithm: NEM theorems and Models,” in The International Joint Conference on Neural Networks (IJCNN). IEEE, 2011, pp. 3178-3183, and 0. Osoba, S. Mitaim, and B. Kosko, “The Noisy Expectation-Maximization Algorithm,” Fluctuation and Noise Letters, vol. 12, no. 13, 1350012-1-1350012-30, September 2013) modifies the EM scheme and achieves faster convergence times on average. The NEM algorithm injects additive noise into the data at each EM iteration. The noise must decay with the iteration count to guarantee convergence to the optimal parameters of the original data model. The additive noise must also satisfy the NEM condition below. The condition guarantees that the NEM parameter estimates will climb faster up the likelihood surface on average.
The NEM Theorem (see O. Osoba, S. Mitaim, and B. Kosko, “Noise Benefits in the Expectation-Maximization Algorithm: NEM theorems and Models,” in The International Joint Conference on Neural Networks (IJCNN) IEEE) states a general sufficient condition when noise speeds up the EM algorithm's convergence to the local optimum of the likelihood surface. The NEM Theorem uses the following notation. The noise random variable N has pdf f(n|o). So the noise N can depend on the observed data O . are the latent variables in the model. {Θ(n)} is a sequence of EM estimates for Θ. Θ* is the converged EM estimate for Θ: Θ*=limn→∞Θ(n). Define the noisy QN function QN(Θ|Θ(n))=EL\O,Θ
Q(Θ*|Θ*)−Q(Θ(n)|Θ*)≧Q(Θ*|Θ*)−QN(Θ(n)|Θ*) (1)
or equivalently
Q
N(Θ(n)|Θ*)≧Q(Θ(n)|Θ*) (2)
holds on average if the following positivity condition holds:
The NEM Theorem states that each iteration of a suitably noisy EM algorithm gives higher likelihood estimates on average than the noiseless EM algorithm gives at each iteration. So the NEM algorithm converges faster than EM does if the data model can be identified. The faster NEM convergence occurs both because the likelihood function has an upper bound and because the NEM algorithm takes larger average steps up the likelihood surface.
Many latent-variable models (such as GMM and HMM) are not identifiable, H. Teicher, “On the mixture of distributions,” The Annals of Mathematical Statistics, pp. 55-73, 1960., H. Teicher, “Identifiability of finite mixtures,” The Annals of Mathematical Statistics, vol. 34, no. 4, pp. 1265-1269, 1963, and thus do not have global likelihood optima. The EM and NEM algorithms converge to local optima in these cases. But the added noise in the NEM algorithm may cause the NEM estimates to search nearby local optima. The NEM Theorem still guarantees that NEM estimates have higher likelihood on average than the EM estimates have for such non-identifiable models.
Gaussian mixture model (GMM) parameter estimation greatly simplifies the NEM positivity condition in (3) O. Osoba, S. Mitaim, and B. Kosko, “Noise Benefits in the Expectation-Maximization Algorithm: NEM theorems and Models,”. Consider the GMM pdf in (??). The model satisfies the positivity condition (3) when the additive noise sample N=(N1, . . . ,ND) for each observation vector o=(o1, . . . , oD) satisfies the following quadratic constraint:
N
d
[N
d−2(μi,k,d−od)]≦0 for all k. (4)
The state sequence S and the Gaussian index Z are the latent variables L for an HMM. The noisy Q-function for the NHMM is
where nt εRD is the noise vector for the observation ot. Then the dth element nt,d of this noise vector satisfies the following positivity constraint:
n
t,d
[n
t,d−2(μi,k,d(n−1)−ot,d)]≦0 for all k (6)
where μi,k(n−1) is the mean estimate at iteration n−1. We also note that noise can provably speed convergence of the Markov chain (see B. Franzke and B. Kosko, “Noise Can Speed Convergence in Markov Chains,” Physical Review E, vol. 84, no. 4, pp. 041112, 2011).
Maximizing the noisy Q-function (5) gives the update equations for the M-step. Only the GMM mean and covariance update equations differ from the noiseless EM because the noise enters the noisy Q-function (5) only through the Gaussian pdf. But the NEM algorithm requires modifying only the covariance update equation (14) because it uses the noiseless mean estimates (13) to check the positivity condition (20). Then the NEM covariance estimate is
NEM technique can apply more generally to other mixture models and exponential family PDFs. O. Osoba and B. Kosko, “The Noisy Expectation-Maximization Algorithm for Multiplicative Noise Injection,” Fluctuation and Noise Letters, vol. 12, no. 13, 1350012-1-1350012-30, September 2013. The next section applies it to the Binomial mixture model.
A Binomial mixture model (BMM) HMM contains a mixture of Binomial PDFs at each HMM state. It is especially popular in bio-informatics and genomics. The PDF fi of an observation o ε ZD+, in a BMM-HMM is
where Pi,k is the parameter of the k-th Binomial distribution at state i of the HMM. We restrict the discussion to non-negative integer noise n for addition to the BMM-HMM since Binomial random variables are non-negative integers.
Next presented is the NEM sufficient condition for a BMM-HMM.
Non-negative integer noise n added to the observation o of a BMM-HMM speeds-up Baum-Welch estimation of its parameters if
Proof. The NEM positivity condition gives
The ratio simplifies into the following component-wise condition for all k:
or equivalently
The Hidden Markov Model Toolkit (HTK) (see S. Young, G. Evermann, D. Kershaw, G. Moore, J. Odell, D. Ollason, V. Valtchev, and P. Woodland, “The HTK book,” Cambridge University Engineering Department, vol. 3, 2002) may be modified to train the NHMM. HTK provides a tool called “HERest” that performs embedded Baum-Welch training for an HMM. This tool first creates a large HMM for each training speech utterance. It concatenates the HMMs for the sub-word units. The Baum-Welch algorithm tunes the parameters of this large HMM.
The NHMM algorithm used (7) to modify covariance matrices in HERest. We sampled from a suitably truncated Gaussian pdf to produce noise that satisfied the NEM positivity condition (6). We used noise variances in {0.001, 0.01, 0.1, 1}. A deterministic annealing factor n−r scaled the noise variance at iteration n. The noise decay rate was τ>0. We used τΣ [1, . . . , 10]. We then added the noise vector to the observations during the update of the covariance matrices (7).
The simulations used the TIMIT speech dataset (see J. S. Garofolo, TIMIT: Acoustic-phonetic Continuous Speech Corpus, Linguistic Data Consortium, 1993. with the standard setup in A. K. Halberstadt and J. R. Glass, “Heterogeneous acoustic measurements for phonetic classification,” in Proc. Eurospeech, 1997, vol. 97, pp. 401-404). The speech signal was parameterized with 12 Mel-Frequency Cepstral Coefficients (MFCC) computed over 20-msec Hamming windows with a 10-msec shift. The first- and second-order finite differences of the MFCC vector were appended with the energies of all three vectors. 3-state left-to-right HMMs were used to model each phoneme with a K-component GMM at each state. K was varied over 1, 4, 8, 16, 32 for the experiments and used two performance metrics to compare NHMM with HMM.
The first metric was the percent reduction in EM iterations for the NHMM to achieve the same per-frame log-likelihood as does the noiseless HMM at iterations 10, 20, and 30. The second metric was the median improvement in per-frame log-likelihood over 30 training iterations.
The learning computer system may include one or more computers at the same or different locations. When at different locations, the computers may be configured to communicate with one another through a wired and/or wireless network communication system.
The learning computer system may include software (e.g., one or more operating systems, device drivers, application programs, and/or communication programs). When software is included, the software includes programming instructions and may include associated data and libraries. When included, the programming instructions are configured to implement one or more algorithms that implement one or more of the functions of the computer system, as recited herein. The description of each function that is performed by each computer system also constitutes a description of the algorithm(s) that performs that function.
The software may be stored on or in one or more non-transitory, tangible storage devices, such as one or more hard disk drives, CDs, DVDs, and/or flash memories. The software may be in source code and/or object code format. Associated data may be stored in any type of volatile and/or non-volatile memory. The software may be loaded into a non-transitory memory and executed by one or more processors. The components, steps, features, objects, benefits, and advantages that have been discussed are merely illustrative. None of them, nor the discussions relating to them, are intended to limit the scope of protection in any way. Numerous other embodiments are also contemplated. These include embodiments that have fewer, additional, and/or different components, steps, features, objects, benefits, and/or advantages. These also include embodiments in which the components and/or steps are arranged and/or ordered differently.
For example, noise is not the only injected perturbation that the invention can use. Other perturbations include chaos, fuzz, and scaled random variables. Nor is adding noise to the data the only form of injection. Multiplicative noise or perturbations will also work as will any other combination of the data with such perturbations. These injections can also reach the probability structure of the model as well as the data.
Unless otherwise stated, all measurements, values, ratings, positions, magnitudes, sizes, and other specifications that are set forth in this specification, including in the claims that follow, are approximate, not exact. They are intended to have a reasonable range that is consistent with the functions to which they relate and with what is customary in the art to which they pertain.
All articles, patents, patent applications, and other publications that have been cited in this disclosure are incorporated herein by reference.
The phrase “means for” when used in a claim is intended to and should be interpreted to embrace the corresponding structures and materials that have been described and their equivalents. Similarly, the phrase “step for” when used in a claim is intended to and should be interpreted to embrace the corresponding acts that have been described and their equivalents. The absence of these phrases from a claim means that the claim is not intended to and should not be interpreted to be limited to these corresponding structures, materials, or acts, or to their equivalents.
The scope of protection is limited solely by the claims that now follow. That scope is intended and should be interpreted to be as broad as is consistent with the ordinary meaning of the language that is used in the claims when interpreted in light of this specification and the prosecution history that follows, except where specific meanings have been set forth, and to encompass all structural and functional equivalents.
Relational terms such as “first” and “second” and the like may be used solely to distinguish one entity or action from another, without necessarily requiring or implying any actual relationship or order between them. The terms “comprises,” “comprising,” and any other variation thereof when used in connection with a list of elements in the specification or claims are intended to indicate that the list is not exclusive and that other elements may be included. Similarly, an element preceded by an “a” or an “an” does not, without further constraints, preclude the existence of additional elements of the identical type.
None of the claims are intended to embrace subject matter that fails to satisfy the requirement of Sections 101, 102, or 103 of the Patent Act, nor should they be interpreted in such a way. Any unintended coverage of such subject matter is hereby disclaimed. Except as just stated in this paragraph, nothing that has been stated or illustrated is intended or should be interpreted to cause a dedication of any component, step, feature, object, benefit, advantage, or equivalent to the public, regardless of whether it is or is not recited in the claims.
The abstract is provided to help the reader quickly ascertain the nature of the technical disclosure. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. In addition, various features in the foregoing detailed description are grouped together in various embodiments to streamline the disclosure. This method of disclosure should not be interpreted as requiring claimed embodiments to require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive subject matter lies in less than all features of a single disclosed embodiment. Thus, the following claims are hereby incorporated into the detailed description, with each claim standing on its own as separately claimed subject matter.
This application is based upon and claims priority to U.S. provisional patent applications 62/025,664, entitled “NOISE-BOOSTED HIDDEN MARKOV MODELS FOR SPEECH RECOGNITION AND OTHER APPLICATIONS,” filed Jul. 17, 2014, attorney docket 094852-0029 and 62/032,385, entitled “NOISE-BOOSTED HIDDEN MARKOV MODELS FOR SPEECH RECOGNITION AND OTHER APPLICATIONS,” filed Aug. 8, 2014, attorney docket 094852-0032. This application is also a continuation-in-part application of U.S. patent application Ser. No. 13/949,048, entitled “ITERATIVE ESTIMATION OF SYSTEM PARAMETERS USING NOISE-LIKE PERTURBATIONS,” filed Jul. 23, 2013, attorney docket 028080-0924, which claimed priority to U.S. provisional patent application 61/674,615, entitled “NOISE-ENHANCED EXPECTATION-MAXIMIZATION ALGORITHM,” filed Jul. 23, 2012, attorney docket 028080-0769. The entire content of each of these applications is incorporated herein by reference.
Number | Date | Country | |
---|---|---|---|
62032385 | Aug 2014 | US | |
61674615 | Jul 2012 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 62025664 | Jul 2014 | US |
Child | 14802760 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 13949048 | Jul 2013 | US |
Child | 62025664 | US |