Method and apparatus for improved probabilistic recognition

Abstract

Probabilistic recognition using clusters and simple probability functions provides improved performance by employing a limited number of clusters each using a relatively large number of simple probability functions.

Description

FIELD OF THE INVENTION

[0003] The present invention is in the field of methods and devices for probabilistic recognition of physical phenomena. More specifically, the present invention is directed to an improved method for speech recognition using a large set of simple probability functions to model speech units grouped into a limited number of clusters.

COPYRIGHT NOTICE

[0004] A portion of the disclosure of this patent document contains material which is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure as it appears in the Patent and Trademark Office patent file or records, but otherwise reserves all copyright rights whatsoever.

BACKGROUND OF THE INVENTION

[0005] This application is filed with a paper appendices of ______ pages which are incorporated as part of this application.

[0006] This invention relates to speech recognition by logic devices or processes, and more particularly to recognizing speech from a large vocabulary using partially-tied probability function mixtures for model state recognition with a reduced number of clusters and an increased number of probability functions used in each cluster.

[0007] This art presumes a basic familiarity with statistics and probability recognition processes, as well as familiarity with the extensive state of the art in recognition systems.

[0008] U.S. Pat. No. 5,825,978, METHOD AND APPARATUS FOR SPEECH RECOGNITION USING OPTIMIZED PARTIAL MIXTURE TYING OF HMM STATE FUNCTIONS provides substantial background information related to the current patent and that patent, and each of its cited references, is incorporated herein by reference. The present invention, however, has applications to other probability recognition paradigms and should not be seen as limited by the above referenced patent.

[0009] Current, state-of-the-art Large-Vocabulary Continuous Speech Recognition (LVCSR) systems are typically based on state-clustered hidden Markov models (HMMs). Typically, these systems use thousands of state clusters, each represented by a Gaussian mixture model with a few tens of Gaussians. These systems use HMMs to model triphone speech units. The number of triphones is usually very large. For example, models with 10,000 triphones are common. Because each triphone is usually modeled by at least three HMM states, this results in about 30,000 HMM states. Each state is typically modeled by a Gaussian mixture model (GMM) with a few Gaussians. Thus, the total number of Gaussian parameters can be on the order of hundreds of thousands. Estimating a separate GMM for each triphone state would require a huge amount of training data. However, because training data is usually limited, it is not possible to reliably estimate such a large number of parameters.

[0010] In one of the first approaches to robust HMM estimation, called the Tied Mixture (TM) HMM, a single set of Gaussian distributions was shared (or tied) across all the states. [1,2] Because the Gaussians were shared, data could be pooled from different HMM states to train the states robustly. Each state was differentiated by a different mixture weight distribution to these shared Gaussians. The shared Gaussians along with the mixture weights defined the state-dependent GMMs. Because of robust parameter estimation, TM HMMs were found to perform significantly better than “fully continuous” HMMs, where each state used a separate GMM.

[0011] To get more detailed models than TM systems, phonetically tied mixture (PTM) systems were proposed. In these systems, a separate Gaussian codebook was shared among all triphone states corresponding to the same base phone. [3]

[0012] A further development in the art was state-clustered HMMs [4,5,6], where the amount of tying was decreased further. This represents the state of the art in speech recognition technology up to the time of the present invention. In this approach, the amount of tying is considerably less than in a TM or PTM system. HMM states are clustered according to acoustic similarity. The states in each cluster either share the same GMM [4,5], or only share the same set of Gaussians but use different mixture weights for each state. [6, 7] A small number of Gaussians is used for each cluster, and improved acoustic resolution is achieved by increasing the number of state clusters.

[0013] In previous work, state-clustered HMMs were experimentally shown to be superior to TM and PTM HMMs (e.g., see [6]). However, in these previous comparisons, the TM and PTM systems had a total of 256 and 4000 Gaussians, respectively—drastically fewer than the total number of Gaussians present in state-clustered system, which had about 24,000 Gaussians. [6] Other previous work with TM and PTM systems [2,8,9] also appears to have used very few Gaussians in comparison to that generally used in state-clustered systems.

[0014] Systems with small numbers of state clusters have previously been studied, but they were not properly explored in that few Gaussians (about 200 to 500) were used in the clusters. This led most practitioners in the art to turn to systems with large numbers of clusters each having few Gaussians.

[0015] What is needed is a speech recognition system or method that has the advantages of conceptually simpler mixture tying systems but gives equal or superior performance to state-clustered systems.

[0016] A further understanding of the invention can be had from the detailed discussion of specific embodiments below. For purposes of clarity, this discussion refers to devices, methods, and concepts in terms of specific examples. However, the method of the present invention may operate with a wide variety of types of recognition and recognition systems. In particular, while parts of the discussion refer to recognition models as hidden Markov models (HMM), it should be understood that HMM can refer to any type of recognition model unless the context requires otherwise. Likewise, while parts of the discussion refer to Gaussians Mixtures as mixtures used to model probability functions, it should be understood that other continuous and discrete or discrete basic probability functions may be used within the context of the invention.

[0017] Furthermore, it is well known in the art that logic systems can include a wide variety of different components and different functions in a modular fashion. Different embodiments of a system can include different mixtures of elements and functions and may group various functions as parts of various elements.

[0018] For purposes of clarity, the invention is described in terms of systems that include many different innovative components and innovative combinations of components. No inference should be taken to limit the invention to combinations containing all of the innovative components listed in any illustrative embodiment in this specification.

[0019] All publications, patents, and patent applications cited herein are hereby incorporated by reference in their entirety for all purposes.

SUMMARY OF THE INVENTION

[0020] In accordance with the invention, speech recognition is provided that takes a new approach to clustering and tying and models used in speech recognition. While it was previously believed that the best performing-systems would be those that used very large numbers of state clusters (about 2000 or more) and few Gaussians per cluster (16 to 32), the current invention uses very few state clusters (about 40 in some embodiments, up to a few hundreds in alternative embodiments) and many Gaussians per state cluster (about 1000). In the present invention, models with far more parameter tying and therefore fewer clusters, like phonetically tied mixture (PTM) models, can give better performance in terms of both recognition accuracy and speed. The present invention can use a conceptually simpler PTM system to achieve faster and more accurate performance than current state-of-the-art state-clustered HMM systems.

[0021] Experimental results have shown between a 5 and 10% improvement in word error rate, while cutting the number of Gaussian distance computations in half, for three different Wall Street Journal (WSJ) test sets, by using a PTM system with 38 phoneclass state clusters, as compared to a state-clustered system with 937 state clusters. For both systems, the total number of Gaussians was fixed at about 30,000.

[0022] The invention will be better understood upon reference to the following detailed description, taken in conjunction with the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

[0023] FIG. 1 is a chart illustrating theoretical state clustering.

[0024] FIG. 2 is a chart illustrating real-world overlap of state clustering.

[0025] FIGS. 3-5 show experimental results for a recognizer according to various embodiments of the invention.

[0026] FIG. 6 illustrates an example logic system which may incorporate the invention according to specific embodiments.

DESCRIPTION OF SPECIFIC EMBODIMENTS

[0027] An understanding of the present invention will be facilitated by a review of the task and art of automatic speech recognition by computers, presented in U.S. Pat. No. 5,825,978, METHOD AND APPARATUS FOR SPEECH RECOGNITION USING OPTIMIZED PARTIAL MIXTURE TYING OF HMM STATE FUNCTIONS, especially FIGS. 1-3 and their related description, incorporated herein by reference.

[0028] According to the present invention, a speech recognizer uses a method for developing and storing model state probability functions that uses fewer clusters with a larger number of Gaussians per cluster than have been generally employed in previous partially-tied systems. Consider a state-clustered system where each state cluster shares the same set of Gaussians, and each triphone state has a separate mixture weight distribution to these shared Gaussians. Suppose one can robustly train at most N state clusters with M Gaussians per cluster, given a certain amount of training data. According to the current invention, by decreasing the number of clusters, while increasing the number of Gaussians per clusters, a system is able to improve the robustness of the Gaussian parameter estimates.

[0029] The advantages of the present invention can be better understood by considering the following. Theoretically, if the Gaussian distributions for the N state clusters do not overlap in acoustic space, then further grouping of the clusters, according to the current invention, will have no effect on performance, as the resulting models will be effectively the same, as shown in FIG. 1.

[0030] However, in reality, state clusters do overlap in acoustic space, as shown in FIG. 2. In the overlap region, Gaussians are separately estimated for each state cluster. This causes two potential problems in prior art systems:

[0031] 1. Because the data in the overlap region is divided between the two state clusters, the Gaussians in this region may not be robustly estimated.

[0032] 2. There may be redundancy between the Gaussians from the two state clusters in the overlap region, resulting in wasted parameters.

[0033] In a specific embodiment, the current invention can address these problems by merging the two clusters into one cluster with a increased number (2M) of Gaussians. Because data from the two clusters is now used to estimate a single set of Gaussians, this method provides more robust estimation in the overlap region. Furthermore, the previously redundant Gaussians can now be more effectively used to increase acoustic resolution, as shown in FIG. 2. According to the current invention, the improved Gaussian estimates and the better acoustic resolution can lead to improved recognition accuracy.

[0034] While merging the two clusters has these advantages, it also has a potential drawback: it may be necessary to separately estimate Gaussians in the overlap regions to be able to aid in discriminating between the clusters, and merging the clusters can reduce the ability to discriminate between clusters. Because decreasing the number of clusters can have both a positive and a negative effect on accuracy, the optimal number of state clusters can, according to the invention, be determined experimentally so as to minimize the word error on development test data.

[0035] State Clustering and Speed

[0036] Computation of the frame-log-likelihoods for all the Gaussian components in each active triphone state during the Viterbi search is a significant cost affecting recognition speed. In SRI's DECIPHER™ speech recognition system, this cost is reduced using techniques of Gaussian caching, Gaussian pruning, and Gaussian shortlists. According to the invention, these methods are affected by changing the number of state clusters, as described below.

[0037] Gaussian Caching

[0038] In Gaussian caching, the log-likelihoods for the Gaussians in a mixture are cached as soon as they are computed for each frame. If the same Gaussian mixture needs to be evaluated at that frame for another triphone state, the cache is used, rather than recomputing the likelihoods of the Gaussians in this mixture. This results in a significant cost saving because many triphone states share the same Gaussian mixture.

[0039] When state clusters are merged, the number of mixtures is reduced, but the number of Gaussians per mixture is increased. Thus, while fewer Gaussian mixtures will be computed and cached, the number of Gaussians in each will be proportionally larger. Therefore, the invention should produce no significant effect due to reducing the number of state clusters on the number of Gaussians computed and cached. However, as discussed in the next section, reducing the number of state clusters can decrease the cost of each Gaussian computation.

[0040] Gaussian Pruning

[0041] When computing the set of Gaussians for a state and frame, it is possible to reduce the amount of Gaussian computations by retaining only those Gaussians whose log-likelihoods are within a threshold of the best Gaussian computed so far. By expanding the diagonal covariance Gaussian likelihood computation, it is easy to see that a system can decide if a Gaussian is within this threshold before computing all the distance components for this frame of speech. This results in a significant reduction in computation cost. Intuitively, the larger the overlap between Gaussians, the larger the number of Gaussians that must be retained for any frame, and the larger the number of distance components that must be computed.

[0042] According to the current invention, when state clusters are merged to create a model with less tying, the redundant Gaussians in the state cluster overlap region are more effectively used to cover the acoustic space of the clusters. The resulting Gaussians will also have smaller variances, as shown in FIG. 2. Because smaller variances imply less Gaussian overlap, according to the invention, the number of Gaussian distance components computed will be reduced.

[0043] Gaussian Shortlists

[0044] Gaussian shortlists are another way to reduce the Gaussian computation during recognition [7]. In this approach, the acoustic space is vector quantized. For each vector quantization (VQ) region, a shortlist of Gaussians that have training data likelihood above some threshold is maintained for each state cluster. During recognition, a system finds the VQ region corresponding to the frame being evaluated, and only computes the likelihoods for the Gaussians in the corresponding shortlists of the state clusters for that VQ region, resulting in a significant speedup.

[0045] When state clusters are merged to create systems with fewer clusters and more tying, according to the current invention, the Gaussian variances are reduced, as in FIG. 2. The reduced variance results in less coverage of the acoustic space by each Gaussian. Thus, Gaussians that previously belonged in a shortlist for a VQ region may no longer have likelihoods high enough to belong in the shortlist for that region. Thus, as a result of the current invention, a reduction in the size of the shortlists is achieved by decreasing die number of state clusters, and a corresponding reduction in Gaussian computation.

[0046] Number of Codebooks

[0047] In one embodiment of the present invention, the number of codebooks is preset to a number such as the number of phones. This provides the benefits of reduced computation as described above and also reduces the processing required for determining clusters.

[0048] According to the present invention, PTM systems, if properly trained, can significantly outperform the currently dominant state-clustered HMM-based approach. In particular, experimental results achieved between 5 and 10% reduction in the word error rate. The number of Gaussians in the shortlists was reduced by half. Finally, at a fixed word error rate, a system according to the invention achieved a factor of 2 reduction in Gaussian distance computation during recognition, and a factor of 5 speedup. The current invention thereby can provide a significant performance gain in accuracy, computation, and speed by using PTM systems as opposed to state-clustered systems.

[0049] In an alternative embodiment, a modified state-clustered system can be used, with a greater number of codebooks, In this embodiment, additional codebooks may be generated using an automatic procedure as described in the '978 patent incorporated herein by reference to identify clusters for sharing mixture components. Again, however, in the present invention the number of codebooks is kept limited while the number of Gaussians per codebook may be expanded so that a total number of Gaussians used in the system remains comparable to that for systems with large numbers of clusters.

[0050] The invention may be used with a training algorithm as described in A. Sankar, “Experiments with a Gaussian Merging-Splitting Algorithm for HMM training for Speech Recognition,” in Proceedings of DARPA Speech Recognition Workshop, (Lansdowne, Va.), February 1998. In such a case, the algorithm as described may produce a variable number of Gaussians per cluster.

EXPERIMENTAL RESULTS

[0051] We experimented using the Wall Street Journal (WSJ) database. For training, we used 18,000 SI284 male training sentences, and for testing we used three different WSJ based test sets. Each test set had 10 speakers, and consisted of about 3600 words, for a total of about 10,900 words. The WSJ domain has been used in previous U.S. Government sponsored speech recognition evaluations. The test sets we used were created for internal development, and are not standardized test sets from the WSJ domain. A 20,000 word bigram language model (LM) was used for recognition. We refer to the three test sets as WSJ1, WSJ2, and WSJ3. We compared two different systems with different levels of tying.

[0052] The first is a state-clustered system with 937 clusters and 32 Gaussians per cluster. We chose this baseline configuration because it has given us good performance in the past. The second is a 38 class PTM system with 789 Gaussians per class. Notice that both systems have a total of about 30,000 Gaussians. Both these systems were trained using the Gaussian Merging Splitting (GMS) algorithm that we recently developed [10]. This method computes only as many Gaussians as can be robustly estimated given the amount of training data, thus giving reliable models. Table 1 compares the word error rates for the two systems on the three different test sets. It is clear that the PTM system is significantly more accurate than the state-clustered system on all three test sets. In particular, the word error rate is reduced by 5 to 10%.

[0053] In Table 1, we did not use Gaussian shortlists. For the remaining experiments, we used Gaussian shortlists and only used the WSJ 1 test set. In Table 2, we compare the size of the Gaussian shortlists for the state-clustered and the PTM systems. Here “size” refers to the number of Gaussians in the shortlists. The number of Gaussians in the PTM system shortlists is half that in the state-clustered shortlists.

TABLE 1
Word error rates for different levels of tying
	Word Error Rate (%)
	System	WSJ1	WSJ2	NABN
	State-clustered	21.65	114.08	18.29
	PTM	20.49	12.58	16.78

[0054]

TABLE 2
Shortlist size for different levels of tying
System          Shortlist Size
State-clustered 5830534
PTM             2773199

[0055] Next, we conducted experiments to evaluate the effect of clustering on recognition computation and speed. We did this by varying the level of pruning in the Viterbi beam search and plotting the word error rate for the WSJ1 test set against different parameters of interest.

[0056] These are the number of Gaussians we start computing per frame, the number of actual distance components computed, and the recognition speed of our system. While the first two parameters are an objective measure of the Gaussian computation cost incurred during recognition, the system speed is implementation-dependent. FIGS. 3, 4, and 5 show these plots. It is clear from these figures that a significant computation saving is gained by using the PTM system over the state-clustered system.

[0057] At a word error rate of 22%, the PTM system has about a factor of 2 less Gaussians started, a factor of 2 less distance component computations, and a factor of 5 speedup. Further, at almost all speeds, the PTM system has a lower word error rate, as shown in FIG. 5. In all three figures we notice that at very high error rates, the PTM system is worse in terms of Gaussian computation and speed (where the curves cross). This occurs because at these error rates, there are only a few active hypotheses in the search per frame, requiring the computation of only a few mixtures. The fact that the state-clustered system has only 32 Gaussians per state cluster as compared to 789 Gaussians for the PTM system then outweighs the computational benefits of the PTM model described in Section 2.2. However, we do not anticipate operating in this high word error region of the curve.

[0058] Variable Clustering

[0059] As described above, the invention uses a much larger number of Gaussians per state cluster than in previous state-clustered systems. However, some phone classes may have very little acoustic variability and thus may need only a few Gaussians for good modeling. For example, the nasal /ng/ is less variable than the unvoiced stop /t/.

[0060] In an alternative embodiment, the invention exploits this fact by developing a per-phone or per-cluster clustering algorithm that determines the number of simple probability functions per phone based on the acoustic variability.

[0061] In one embodiment, to measure a phone or cluster's variability, the invention agglomeratively clusters the HMM states for each phone, using a weighted-by-counts entropy distance between the mixture weight distributions of each state, as is known in the art. Clustering is stopped when the average distance reaches a pre-specified relative threshold. The number of resulting state clusters is a measure of a phone's acoustic variability. In one tested implementation, the number of Gaussians for a phone was proportional to the acoustic variability, with a pre-set minimum and maximum number of Gaussians.

[0062] Table 3 shows the word error rate on the 1996 H4 PE development test set and the number of Gaussians for three different female models trained on the first 100 hours of H4 training data. These recognition runs used a 48,000-word bigram LM.

[0063] 1997-eval is a state-clustered model used for the 1997 DARPA H4 evaluations. PTM-1788 is a PTM model with 1788 Gaussians per phone class, and Clustered-PTM is a model created by applying the per-phone Gaussian clustering algorithm to PTM-1788. From the table, it can be seen that the PTM and state-clustered systems gave the same word error rate. A factor of 5 reduction in the number of Gaussians was achieved using the per-phone Gaussian clustered PTM model, with no difference in the word error rate. The drastic reduction in Gaussians also decreases the amount of computation during recognition.

TABLE 3
Word error rates and number of Gaussians for different Models
	Model	Word (%) Error	Number of Gaussians
	1997-eval	39.4	67,200
	PTM-1788	39.7	69,732
	Clustered-PTM	39.3	12,758

[0064] Embodiment in a Programmed Digital Apparatus

[0065] The invention may be embodied in a fixed media or transmissible program component containing logic instructions and/or data that when loaded into an appropriately configured computing device cause that device to perform recognition according to the invention.

[0066] FIG. 6 shows digital device 700 that may be understood as a logical apparatus that can read instructions from media 717 and/or network port 719. Apparatus 700 can thereafter use those instructions to perform recognition tasks. One type of logical apparatus that may embody the invention is a computer system as illustrated in 700, containing CPU 707, optional input devices 709 and 711, disk drives 715 and optional monitor 705. Fixed media 717 may be used to program such a system and could represent a disk-type optical or magnetic media or a memory. Communication port 719 may also be used to program such a system and could represent any type of communication connection.

[0067] The invention also may be embodied within the circuitry of an application specific integrated circuit (ASIC) or a programmable logic device (PLD). In such a case, the invention may be embodied in a computer understandable descriptor language which may be used to create an ASIC or PLD that operates as herein described.

[0068] The invention also may be embodied within the circuitry or logic processes of other digital apparatus, such as telephones, commications cameras, displays, image editing equipment, etc.

[0069] The invention is explained herein with regard to specific and alternative embodiments. Variations on these embodiments and other embodiments will be apparent to those of skill in the art. The invention therefore should not be limited except as provided in the attached claims.

[0070] It is understood that the examples and embodiments described herein are for illustrative purposes only and that various modifications or changes in light thereof will be suggested to persons skilled in the art and are to be included within the spirit and purview of this application and scope of the appended claims. All publications, patents, and patent applications cited herein are hereby incorporated by reference in their entirety for all purposes.

REFERENCES

[0071] 1. X. Huang and M. Jack, “Semi-Continuous Hidden Markov Models for Speech Signals,” Computer Speech and Language, vol. 3, pp. 239-252, 1989.

[0072] 2. J. Bellagarda and D. Nahamoo, “Tied Mixture Continuous Parameter Modeling for Speech Recognition,” IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 38, pp. 2033-2045, December 1990.

[0073] 3. D. Paul, “The Lincoln Robust Continuous Speech Recognizer,” in Proceedings IEEE International Conference on Acoustics, Speech, and Signal Processing, pp. 449-452, May 1989.

[0074] 4. M. Y. Hwang, X. Huang, and F. Alleva, “Predicting Unseen Triphones with Senones,” in Proceedings IEEE International Conference on Acoustics, Speech, and Signal Processing, pp. II-311-II-314, 1993.

[0075] 5. P. Woodland, J. Odell, V. Valtchev, and S. Young, “Large Vocabulary Continuous Speech Recognition Using HTK,” in Proceedings IEEE International Conference on Acoustics, Speech, and Signal Processing, pp. II-125-II-128, 1994.

[0076] 6. V. Digalakis and H. Murveit, “High Accuracy Large Vocabulary Speech Recognition Using Mixture Tying and Consistency Modeling,” in Proceedings of the DARPA Human Language Technology Workshop, (Plainsboro, N.J.), 1994.

[0077] 7. V. Digalakis, P. Monaco, and H. Murveit, “Genones: Generalized Mixture Tying in Continuous Hidden Markov Model Based Speech Recognizers,” IEEE Transactions on Speech and Audio Processing, vol. 4, no. 4, pp. 281-289, 1996.

[0078] 8. O. Kimball and M. Ostendorf, “On the Use of Tied-Mixture Distributions,” in Proceedings of the DARPA Human Language Technology Workshop, (Plainsboro, N.J.), 1993.

[0079] 9. D. B. Paul, “The Lincoln Large Vocabulary Stack-Decoder Based HMM CSR,” in Proceedings of the DARPA Human Language Technology Workshop, pp. 399-404, 1994.

[0080] 10. A. Sankar, “Experiments with a Gaussian Merging-Splitting Algorithm for HMM training for Speech Recognition,” in Proceedings of DARPA Speech Recognition Workshop, (Lansdowne, Va.), February 1998.

[0081] 11. A. Sankar, R. Gadde and F. Weng, “SRI's 1998 Broadcast News System—Toward Faster, Better, Smaller Speech Recognition” in Proceedings of the DARPA Broadcast News Workshop, (Washington, D.C.), February 1999.

[0082] 12. A. Sankar, R. Gadde. “Parameter Tying and Gaussian Cluster for Faster, Better, and Smaller Speech Recognition, in Proceedings of EUROSPEECH (Budapest, Hungary), September 1999.

Claims

1. In a speech recognition system using a method for recognizing human speech, the method being of the type comprising the steps of: selecting a model to represent a selected subunit of speech, the model having associated with it a plurality of states and each state having associated with it a probability function, the probability function having undetermined parameters, the probability functions being represented by a mixture of simple probability functions; extracting features from a set of speech training data; using the features to determine parameters for the probability functions in the model, an improved method for recognizing speech, the improvement comprising the steps of: determining states that may be represented by a set of simple probability functions; clustering said states that may be represented by a set of simple probability functions into a limited number of clusters; defining a plurality of cluster codebooks, one cluster codebook associated with each one of said clusters and said cluster codebooks having a large number of simple probability functions per code book; and estimating the simple probability functions in each cluster codebook to fit states in that cluster.

2. The method according to claim 1 wherein the number of clusters is kept low to improve processing speed while the number of simple probability functions per cluster is increased for greater recognition accuracy.

3. The method according to claim 1 wherein the number of clusters is between approximately 40 and 200 and wherein at least one cluster has assigned to it 500 to 2000 simple probability functions.

4. The method according to claim 1 wherein the number of clusters is more than 10 and wherein the ratio of the number of clusters to the total number of simple probability functions in the system is less than 0.002.

5. The method according to claim 1 wherein the number of clusters is more than 9 and at least one cluster has more than approximately 1,000 simple probability functions.

6. The method according to claim 1 wherein the simple probability functions are Gaussians.

7. The method according to claim 1 wherein different numbers of simple probability functions are used in different clusters.

8. The method according to claim 7 wherein the number of simple probability functions used for a particular cluster is determined by a training algorithm.

9. The method according to claim 7 wherein the number of simple probability functions used for a particular cluster is indicated by a human system designer.

10. The method according to claim 1 wherein the number of said clusters is equal to the number of phones in the system.

11. The method according to claim 1 wherein the model is a three-state Hidden Markov Model.

12. The method according to claim 1 wherein states are clustered according to an agglomerative hierarchical clustering scheme.

13. The method according to claim 1 wherein states are clustered so as to nearly eliminate overlap between clusters.

14. The method according to claim 1 further comprising: caching log-likelihoods for the simple probability functions in a mixture as soon as they are computed for a frame so that if the same mixture needs to be evaluated at that frame for another triphone state, the cache is used.

15. The method according to claim 1 wherein redundant simple probability functions in the state cluster overlap region are more effectively used to cover the acoustic space of the clusters, resulting in smaller variances and a reducing the number of distance components to be computed.

16. The method according to claim 1 further comprising: reducing the size of a simple probability function shortlists by decreasing the number of state clusters with a corresponding reduction in simple probability function computations.

17. A computer readable data file stored or transmitted on a media that when loaded into an appropriately configured computer system will enable the system to perform probabilistic recognition using a set of probabilistic models including a limited number of state clusters, at least one state clustered represented by a large number of simple probability functions.

18. A speech recognizer comprising: a logic processing device; storage means; a set of probabilistic models stored in the storage means; said models including a limited number of state clusters, at least one state clustered represented by a large number of simple probability functions; a feature extractor in the computer for extracting feature data capable of being processed by said computer from a speech signal; and recognizing means for matching features from unidentified speech data to the models to produce a most likely path through the models where the path defines the most likely subunits and words in the speech data.

19. A method for developing tied-transform HMMs able to estimate probability function parameters with little data comprising: selecting a small enough number of state clusters M so that robust estimates are possible; training an HMM for a smaller number of state clusters M, for which it is assumed there is enough data to robustly estimate each probability functions parameters; determining for each state cluster in the larger HMM an ancestor descendent relationship with a cluster in the smaller HMM; defining a mapping from the smaller to the larger HMM in terms of the ancestor descendent relationship; wherein probability functions in the state clusters of the larger HMM are transformed versions of the ancestor probability functions in the smaller HMM