CROSS-REFERENCE TO RELATED APPLICATION
BACKGROUND OF THE INVENTION
This application is a continuation-in-part of U.S. application Ser. No. 10/127,184, entitled, “Pattern Matching for Large Vocabulary Speech Recognition Systems, filed Apr. 22, 2002.
The present invention relates generally to large vocabulary continuous speech recognition system, and more particularly, to a method for improving pattern matching in a large vocabulary continuous speech recognition system.
Pattern matching is one of the more computationally intensive aspect of the speech recognition process. Conventional pattern matching involves computing similarity measures for each acoustic feature vector in relation to each of the acoustic models. However, due to the large number of acoustic models, only a subset of acoustic models may be loaded into the available memory at any given time. In order to compute similarity measures for a given acoustic feature vector, conventional pattern matching requires a number of I/O operations to load and unload each of the acoustic models into the available memory space.
- SUMMARY OF THE INVENTION
Therefore, it is desirable to provide an improved method of pattern matching that reduces the number I/O operations associated with loading and unloading each acoustic model into memory.
In accordance with the present invention, a method is provided for improving pattern matching in a speech recognition system having a plurality of acoustic models. The improved method includes: receiving continuous speech input; generating a sequence of acoustic feature vectors that represent temporal and spectral behavior of the speech input; loading a first group of acoustic feature vectors from the sequence of acoustic feature vectors into a memory workspace accessible to a processor; loading an acoustic model from the plurality of acoustic models into the memory workspace; and determining a similarity measure for each acoustic feature vector of the first group of acoustic feature vectors in relation to the acoustic model. Prior to retrieving another group of acoustic feature vectors, similarity measures are computed for the first group of acoustic feature vectors in relation to each of the acoustic models employed by the speech recognition system. In this way, the improved method reduces the number I/O operations associated with loading and unloading each acoustic model into memory.
In accordance with another aspect of the invention, a method is provided for processing speech data utilizing high speed cache memory. The cache memory has an associated cache mechanism for transfer of data from system memory into cache memory that may operate automatically or under program control, depending on the features provided by the processor. First, a main table of speech data in system memory is provided along with a list that establishes a processing order of a subset of said speech data. In this regard, the tem “list” is intended to encompass any data structure that can represent sequential information (such as the sequential information found in a speech utterance).
The method involves copying the subset of said speech data into a sub-table that is processed such that entries in said sub-table occupy contiguous memory locations. Then the sub-table is operated upon using a speech processing algorithm, and the cache mechanism associated with said high speed cache memory is employed (automatically or programmatically) to transfer the sub-table into said high speed cache memory. In this way, the speech processing algorithm accesses the subset of speech data at cache memory access rates and thereby provides significant speed improvement.
BRIEF DESCRIPTION OF THE DRAWINGS
For a more complete understanding of the invention, its objects and advantages, reference may be had to the following specification and to the accompanying drawings.
FIG. 1 is a block diagram depicting an exemplary speech recognition system;
FIG. 2 is a flowchart illustrating a method for improving pattern matching in large vocabulary speech recognition systems in accordance with the present invention;
FIG. 3 is a block diagram illustrating how the improved method of pattern matching may be distributed across multiple processing nodes in accordance with the present invention;
FIGS. 4A-4C are diagrams illustrating how the decoding processing may be distributed amongst various processing nodes in accordance with the present invention;
FIG. 5 is a diagram depicting an exemplary lexical search space;
FIGS. 6 and 7 are block diagrams depicting distributed architectural arrangements for large vocabulary speech recognition systems in accordance with the present invention;
FIG. 8 is a block diagram illustrating a general relationship between main system memory and high speed cache memory;
FIG. 9 is an algorithm block diagram illustrating a preferred embodiment of the packed distribution and localized trellis access method;
FIG. 10 is an exemplary word graph illustrating how a processing order may be extracted from a spoken utterance;
FIG. 11 is a data flow diagram providing an example of how the packed mixture table (sub-table) is populated using the packed distribution and localized trellis access technique; and
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
FIG. 12 is a flowchart describing a presently preferred method for implementing the packed distribution and localized trellis access technique.
FIG. 1 illustrates an exemplary speech recognition system. The system operates in two phases: a training phase, during which the system learns the reference patterns representing the different speech sounds (e.g., phrases, words, phones) that constitute the vocabulary of the application; and a recognition phase, during which an unknown input pattern is identified by considering the set of references. During the training phase, each reference is learned from spoken examples and stored either in the form of templates obtained by some averaging method (in template-matching systems) or acoustic models that characterize the statistical properties of patterns (like in stochastic systems). One of the most popular stochastic systems utilizes a statistical modeling approach employing Hidden Markov Models (HMM).
The exemplary speech recognizer performs the recognition process in three steps as shown in FIG. 1. First, speech analysis and feature extraction 10 is performed on the input speech. This step generates a sequence of acoustic feature vectors representing the temporal and spectral behavior of the speech input. In general, an input speech signal is partitioned into a sequence of time segments or frames. Spectral features are then extracted from each frame using a variety of well known techniques.
Next, acoustic pattern matching occurs at step 12. During this step, a similarity measure is computed between each frame of input speech and each reference pattern. The process defines a local measure of closeness between acoustic feature vectors and further involves aligning two speech patterns which may differ in duration and rate of speaking. The pattern classification step uses a plurality of acoustic models 14 generated during the training phase.
A diagram of a simple Hidden Markov Model is shown at 20 of FIG. 1. As noted above, Hidden Markov Models are commonly employed as acoustic models by speech recognition systems. For illustration purposes, a three-state Hidden Markov Model is depicted having the states designated s1, s2 and s3. It is readily understood that HHMs could employ a different number of states. Moreover, it is understood that the present invention is not limited to HMMs, but is applicable to speech recognition systems employing other types of acoustic models.
Each Hidden Markov Model includes a collection of probabilities associated with the states themselves and transition amongst the states. Because probability values associated with each state may be more complex than a single value could represent, some systems will represent probability in terms of a Gaussian distribution. To provide a more robust model, a mixture of Gaussian distributions may be used in a blended manner to represent probability values as shown diagrammatically at 26 and referenced by a mixture index pointer 28. Thus, associated with each state is a mixture index pointer which in turn identifies the Gaussian mixture density data for that state.
Transitions amongst the states are illustrated by arrows. Each self-loop transition has an associated transition probability as depicted at 22; whereas each transition to another state also has an associated transition probability as depicted at 24. Likewise, transition probabilities may be represented by Gaussian distributions data or Gaussian mixture density data.
In the context of large vocabulary speech recognizers, Hidden Markov Models are typically used to model sub-word units, such as phonemes. However, speech recognition systems that employ word-level acoustic models or acoustic models based on another speech sub-component are also within the scope of the present invention. For more information regarding the basic structure of Hidden Markov Modeling, see Junqua, Jean-Claude and Haton, Jean-Paul, Robustness in Automatic Speech Recognition, Fundamentals and Applications, Kluwer Academic Publishers, 1996.
Speech recognition concludes with a decoding step 16. The probability that a particular phoneme was spoken is provided by the acoustic models as part of the pattern matching process. A sequence of words can then be constructed by concatenating the phonemes observed during the pattern matching process. The process of combining probabilities for each possible path and searching through the possible paths to select the one with highest probability is commonly referred to as decoding or searching. In other words, the decoding process selects a sequence of words having the highest probability given the observed input speech. A variety of well known searching algorithms may be used to implement the decoding process.
In one aspect of the present invention, an improved method is provided for performing pattern matching in a large vocabulary continuous speech recognition system as shown in FIG. 2. Rather than determine similarity measures for each acoustic feature vector as it is received, a group of acoustic feature vectors are buffered into cache memory that is accessible to a data processor. Similarity measures are then determined for each acoustic feature vector in the group of vectors. This improved method may be herein referred to as “horizontal caching”.
Referring to FIG. 2, a first group of acoustic feature vectors is retrieved into a memory workspace at step 32. Similarly, one or more acoustic models are also loaded into the memory workspace at step 34, where the number of acoustic models loaded into memory is a subset of the acoustic models employed by the speech recognition system. In the case of HMMs, the Gaussian distribution data or Gaussian mixture density data which serves as the basis for the acoustic model is loaded into memory. One skilled in the art will readily recognize that the term “memory workspace” preferably refers to cache memory or some other data store readily accessible to the data processor. It is envisioned that the number of acoustic feature vectors associated with the first group and the number of acoustic models loaded into the memory workspace should be selected to optimize use of the available memory space.
A similarity measure can then be computed at step 36 for each acoustic feature vector in the first group of vectors. For example, a Gaussian computation may be performed for each acoustic feature vector as is well known in the art. Resulting similarity measures may be stored in an output memory space which is also accessible to the processor performing the computations. By performing the similarity computation for a group of acoustic feature vectors, the present invention reduces the number I/O operations required to load and unload each acoustic model.
Prior to retrieving additional acoustic models, the acoustic models currently resident in the memory workspace are removed at step 38. Additional acoustic models are then loaded into the memory space at step 42. If desired the removal step 38 can be performed concurrently with the loading step 42; the loading step can overwrite what is already stored in the memory workspace, thereby removing the models then resident. Similarity measures are computed for each acoustic feature vector in the first vector group in relation to each of the additional acoustic models resident in the memory workspace at step 36. Again, the resulting similarity measures may be stored in an output memory space which is also accessible to the processor performing the computations. This process is repeated via step 40 until similarity measures are computed for the first group of acoustic feature vectors in relation to each of the acoustic models employed by the speech recognition system.
Once similarity measures have been determined for the first group of acoustic feature vectors, the search process is performed at step 44. In particular, the search process updates the search space based on the similarity measures for the first group of acoustic feature vectors. It is to be understood that this aspect of the present invention is not limited to a particular searching algorithm, but may be implemented using a variety of well known searching algorithms.
Contemporaneous with the search process, a subsequent group of acoustic feature vectors may be retrieved into the memory workspace at step 48. A similarity measure is computed for each acoustic feature vector in this subsequent group as described above. In other words, acoustic models are loaded and unloaded into the memory workspace and a Gaussian computation is performed for each acoustic feature vector in relation to the acoustic models resident in the memory workspace. This process is repeated via step 40 until similarity measures are computed for the subsequent group of acoustic feature vectors in relation to each of the acoustic models employed by the speech recognition system. It is envisioned that the first group of acoustic feature vectors is removed from the memory workspace prior to loading the subsequent group of acoustic feature vectors into the memory workspace. One skilled in the art will readily recognize that this is an iterative process that is performed for each of the acoustic feature vectors that represents the input speech.
It is further envisioned that the improved method for performing pattern matching may be distributed across multiple processing nodes as shown in FIG. 3. Rather than performing the pattern matching process on a single data processor, the process is partitioned among a plurality of processing nodes. As will be further described below, each processing node is responsible for computing similarity measures for a particular acoustic model or group of acoustic models.
An acoustic front-end node 52 is receptive of speech input and operable to generate a sequence of acoustic feature vectors as is known in the art. The acoustic front-end node 52 is further able to replicate the sequence of acoustic feature vectors 54 and distribute the replicated sequences 54 amongst the plurality of pattern matching nodes 56. It is envisioned that the replicated sequence of acoustic feature vectors may be partitioned into groups of vectors which are periodically or upon request communicated to the plurality of pattern matching nodes.
Each pattern matching node 56 is comprised of a data processor 58 and a memory space 59 accessible to the data processor 58. To perform pattern matching, each pattern matching node 56 is adapted to receive the replicated sequence of acoustic feature vectors 54 from the acoustic front-end node 52. As described above, each pattern matching node 56 is operable to load one or more acoustic models 60 into a resident memory space, and then determine similarity measures for each acoustic feature vector in relation the loaded acoustic models. In this approach, each pattern matching node 56 is responsible for a predetermined range of acoustic models, such that computation of similarity measures for a given acoustic feature vector or group of vectors can occur in parallel, thereby further improving the overall performance of the speech recognition process.
In another aspect of the present invention, the decoding process may be distributed amongst a plurality of processing nodes. In general, the search space is comprised of observed acoustic data (also referred to as the potential search space). Referring to FIG. 4A, the search space may be diagrammatically represented as a plurality of nodes 62, where each node signifies a state of a certain phoneme of a certain word of a certain word history for language model conditioning. The states of all phonemes of all of the words comprise the search space. The search space may be further segmented to include a potential search space and an active search space. The active search space is the area being explored by a search algorithm at a given time. In contrast, the potential search space is defined as the maximum possible active search space. In FIG. 4A, the black nodes indicate the active search space; whereas all of the nodes comprise the potential search space.
To further reduce computational processing, the observed acoustic data may be partitioned amongst a plurality of processing nodes as shown in FIG. 4B. A searching operation is then performed on the observed acoustic data allocated to each processing node, such that at least some of the searching operations occur concurrently on different processing nodes. Although a Viterbi searching algorithm is presently preferred, it is readily understood that other known search algorithms, such as a stack decoding algorithm, a multi-pass search algorithm or a forward-backward search algorithm, are within the scope of the present invention.
Partitioning the observed acoustic data further includes defining link data 64 that is indicative of the relationships between the segmented acoustic data residing at the different processing nodes. Since each processing node only evaluates a subset of the observed acoustic data, link data is maintained at each of the processing nodes. As further describe below, changes in the link data is communicated amongst the plurality of processing nodes.
In FIG. 4B, the search space is segmented in a manner that minimizes the number of required links amongst the segmented acoustic data. However, this division does not maximize the available processing power. The searching operation associated with a third processing node 68 is dependent upon the completion of the searching operations associated with a first processing node 66 and a second processing node 67. Alternatively, the search space may be partitioned as shown in FIG. 4C. In this case, each state of a certain phoneme is sequentially assigned to a different processing node. Although this exemplary division provides a better utilization of the available processing power, it also requires a considerable amount of link data. Similarly, it is also envisioned that the observed acoustic data may be allocated in proportion to the processing power associated with each processing node From such discussions, it is readily understood that the search space may be partitioned in accordance with a predefined criteria including but not limited to the criteria (or combinations thereof) discussed above.
For illustration purposes, a decoding process based on lexical trees is further described below. Lexical trees generally represent the pronunciations of words in the vocabulary and may be constructed by concatenating the phonemes observed during the pattern matching process. Each node in a lexical tree is associated to a state of a certain phoneme of a certain word of a certain word history for language model conditioning. The states of all phonemes of all words have been compiled into lexical trees. These trees are replicated for word history language model conditioning.
Referring to FIG. 5, the search space 70 is comprised of a plurality of lexical trees 72. In this case, one or more lexical trees may be allocated to a particular processing node. A terminating node in a lexical tree signifies a unique word in the lexicon. Links 74 are used to interconnect selective terminating nodes of different lexical trees 72, thereby forming likely word sequences. To derive at the most likely word sequence, a searching algorithm, such as the Viterbi searching algorithm, is used to traverse the lexical trees in a manner that is well known in the art.
FIG. 6 illustrates an architecture that further distributes the speech recognition process across multiple processing nodes in accordance with the present invention. The distributed architecture 80 is comprised of a pattern matching subsystem 82 and a lexical searching subsystem 84 interconnected by a communication link 86.
The pattern matching subsystem 82 is comprised of a plurality of pattern matching nodes 88. To perform pattern matching, each pattern matching node 88 is adapted to receive a replicated sequence of acoustic feature vectors from an acoustic front-end node (not shown). As described above, each pattern matching node 88 determines similarity measures for a predetermined range of acoustic models, such that computation of similarity measures for a given acoustic feature vector occurs in parallel. Resulting similarity measures are then communicated from each of the pattern matching nodes 88 via the communication link 86 to the lexical search subsystem 84.
Resulting similarity measures are preferably communicated in a multicast mode over an unreliable link. A reliable link typically require a connection protocol, such as TCP, which guarantees that the information is received by the intended recipient. Reliable links are typically more expensive in term of bandwidth and latency, and thus should only be used when data needs to be received. In contrast, an unreliable link usually does not require a connection to be opened but does not guarantee that all transmitted data is received by the recipient. In an exemplary embodiment, the communication link 86 is a standard Ethernet link (e.g., 100 Mbits/sec). Although an unreliable link is presently preferred to maximize throughout, a reliable link may also be used to communicate similarity measures between the pattern matching subsystem and the lexical searching subsystem.
Similarly, the lexical search subsystem 84 is comprised of a plurality of searching nodes 90. The search space is partitioned such that each searching node 90 is responsible for evaluating one or more of the lexical trees which define the search space. To do so, each searching node 90 is adapted to receive similarity measures from each of the pattern matching nodes 88 in the pattern matching subsystem 82.
If a searching node does not receive some of the similarity measure data that it needs, the node could either compute it or ask for it to be retransmitted. To recompute similarity measures, the searching node would need to access to all of the acoustic models which could constitute a considerable memory use. On the other hand, retransmitting similarity measures is equivalent to implementing reliable multicast. Although the approach is expensive in terms of bandwidth and especially in terms of latency, it may be feasible in some applications.
For instance, the latency problem due to retransmissions inherent with the reliable multicast mode may not be a problem in the horizontal caching technique described above. To maximize throughput on the communication link, assume that a daisy chain is constructed with reliable links between the pattern matching nodes 88. The daisy chain is used to synchronize the transmission of the similarity measures using a round-robin approach. This approach has the advantage that the pattern matching nodes would not try to write on the shared link at the same time, thereby creating collisions and possible retransmissions.
Using this approach, the first pattern matching node would write the first 10 frames (equivalent to 100 milliseconds of speech) of its output cache on the shared non-reliable link. The first node then signals the next node on the chain that it is now its turn to transmit data. The next node will transmit its data and then signal yet another node. Assuming 8 pattern matching nodes, the total amount of data each node will have to send over the shared medium is 10 frames×10 kminutes/8 nodes×4 bytes=50 Kbytes=0.4 Mbits. To complete this process for 8 nodes, it takes 32 milliseconds over a 100 Mbits per second shared link, not accounting for overhead, latency due to the transmission and synchronization of the daisy chain. Since only one third of the total aggregate bandwidth of the communication link has been used, the remainder of the bandwidth could be used for retransmission associated with the reliable multicast. One skilled in the art will readily recognize that if the latencies are too high, the horizontal caching technique provides the flexibility to increase the batch size to more than 10 frames, therefore reducing the sensitivity to latencies.
Each searching node 90 only processes a subset of the lexical trees in the search space. To do so, each searching node 90 needs to know the state of its associated lexical trees as well as data indicating the links between all of the lexical trees in the search space. Thus, each searching node further includes a data store for maintaining the link data.
Since processing of associated lexical trees by a searching node may result in changes to the link data, each searching node 90 is further operable to communicate changes to the link data to each of the other searching nodes in the lexical search subsystem. Here, the communication problem is more difficult because synchronization up to the frame time (e.g., 10 milliseconds) and reliability must be guaranteed. Although a shared communication link may be feasible, a switching network is preferably used to link searching node in the lexical search subsystem. In particular, each searching node 80 is interconnected by a switching fabric 92 having a dedicated link.
In operation, each searching node 90 will be listening and reading the similarity measures from the pattern matching subsystem 82. In this case, each searching node 90 is multi-threaded, so that reading from the communication link can be done in parallel with processing of lexical trees. At the end of each frame, each search node 90 will send the likely word endings and a few other statistics (e.g., likelihoods histograms used to adapt the beam search) to a search reduction server 94. The search reduction server 94 is operable to combine information about word endings, apply a language model to generate a new (global) search state and sent the search state back (in multicast mode) to each searching node 90. All of this process has to be accomplished in a time window smaller that the frame rate, and in a reliable way, since the search state has to be maintained consistent across all nodes. Therefore, efficient reliable multicast is preferably employed. In addition, the search reduction server is further operable to generate the recognized sentence and to compute statistics, like the confidence measure or the speaker id, as post processing.
FIG. 7 illustrates an alternative distributed architecture were the searching nodes 90 are directly linked between with a shared medium 98. Assuming that each searching node 90 is independently performing the search reduction processes in a distributed way, there is no need for a search reduction server. However, each node will have to store the language model and employ an N to N reliable multicast communication mode. This solution may be less expensive but more difficult to implement.
Reducing the size of the search space is another known technique for reducing computational processing associated with the decoding processing. Histogram pruning is one known technique for reducing the number of active nodes residing in the search space. One known technique for achieving N best (or approximately N best) pruning is through the computation of a histogram. The histogram represents the probability density function of the scores of the nodes. It is defined as y=f(X), where X is the score and y is the number of nodes a given time t with the score. Since scores are real numbers, X does not represent a specific value, but rather a range.
For illustration purposes, a simplistic example of histogram pruning is provided below. Suppose we have 10 active states at time t, and that we should wish to retain only 5 of them. Assume the active states are as follows:
- s0: score 3 associated to node n0
- s1: score 2 associated to node n1
- s2: score 5 associated to node n2
- s3: score 4 associated to node n3
- s4: score 4 associated to node n4
- s5: score 3 associated to node n5
- s6: score 5 associated to node n6
- s7: score 3 associated to node n7
- s8: score 2 associated to node n8
- s9: score 5 associated to node n9
Thus, the histogram maps:
- f(2)=2 (states s1, and s8)
- f(3)=3 (states s0, s5, s7)
- f(4)=2 (states s3 and s4)
- f(5)=3 (states s2, s6, s9)
We do not need to know which states are associated with which value of X, and therefore a simple array y=f(X) is sufficient.
Next, to identify the N=5 best, we just look at the histogram to compute the threshold, T, corresponding to the pruning. If T=6 or above, no states satisfy score(s)>=T. If T=5, then add backwards the number of nodes s which satisfy score(s)>=T: f(5)=3. In this case, only three node meet the threshold. Since three nodes is insufficient to meet our pruning criteria (3<N=5), then we continue by setting T=4. In this case, five nodes meet the threshold. The threshold (T=4), can then be applied to the list of nodes as follows:
- s0: score 3 associated to node n0===>remove
- s1: score 2 associated to node n1===>remove
- s2: score 5 associated to node n2===>KEEP
- s3: score 4 associated to node n3===>KEEP
- s4: score 4 associated to node n4===>KEEP
- s5: score 3 associated to node n5===>remove
- s6: score 5 associated to node n6===>KEEP
- s7: score 3 associated to node n7===>remove
- s8: score 2 associated to node n8===>remove
- s9: score 5 associated to node n9===>KEEP
Histogram pruning may be implemented in the distributed environment of the present invention as described below. Assume the search space is divided amongst three search nodes, K1, K2, and K3, such that:
- s0: score 3: processed by node K1
- s1: score 2: processed by node K2
- s2: score 5: processed by node K3
- s3: score 4: processed by node K1
- s4: score 4: processed by node K1
- s5: score 3: processed by node K1
- s6: score 5: processed by node K2
- s7: score 3: processed by node K2
- s8: score 2: processed by node K3
- s9: score 5: processed by node K3
To identify 5 active states, each search processing node computes its own histogram as follows:
- K1: f(3)=2 (s0 and s5), f(4)=2 (s3 and s4)
- K2: f(2)=1 (s1), f(3)=1 (s6), f(5)=1 (s6)
- K3: f(2)=1 (s8), f(5)=2 (s2,s9)
Unfortunately, this example, is not very exemplary of the distribution of scores. The distribution is typically in an identifiable form, such as exponential. In other words, y=f(M−X)=alpha*exp(1/alpha*(M−X)). In this case, the threshold may be computed from estimations for the parameters alpha and M. Specifically, the threshold is T=M−1/alpha*log N, where M is the maximum score and the expectation (average value) is M−1/alpha.
To compute the threshold, an algorithm is implemented at each searching node. The algorithm involves looping through all the nodes and computing the mean value and max value of all scores. Let Mk denote the max score on search processing node Kk, Ek denote the mean value of the scores on node Kk, and Wk be the number of active nodes on Kk, where k=1, 2 . . . n.
The overall threshold may be recovered by using Mk, Ek, and Wk from each of the searching nodes. The overall maximum M is equal to the largest Mk and the overall mean is 1/(sum Wk)*(sum of Wk*Ek). Since Mk, Ek, and Wk are the only entities that need to be transmitted, they are called sufficient statistics for the computation of the threshold T. Furthermore, these statistics are much smaller than the large array y=f(X).
- Packed Distribution and Localized Trellis Access
Based on these sufficient statistics, computation of a threshold is done at one of the processing nodes (possibly the root node) and then transmitted back to each of the search nodes. The threshold is applied to the active nodes at each processing node as previously explained.
Large vocabulary speech applications will typically employ a very large number of speech parameters. For example, an exemplary large vocabulary speech recognition system may require a Gaussian mixture table containing 100,000 Gaussians, or more. There is a class of speech processing problems that initially require access to the entire table, but that later constrain access to a subset of the entire table. For example, in a multi-pass recognizer, the speech processing algorithm uses the first pass to constrain the search space used by subsequent passes.
The need to deal with massive amounts of data makes large vocabulary speech applications highly processor intensive. Unfortunately, conventional processing algorithms do little to combat this problem, but instead place the computational burden on comparatively expensive processors. This traditional “brute force” approach has placed large vocabulary applications off limits for a variety of consumer products that do not have powerful processors. However, as will be more fully explained herein, it is possible to significantly improve processing throughput, and to significantly reduce processor overhead, by taking advantage of the a priori knowledge of the temporal order or spoken order inherent in many speech applications. As will be more fully explained, these improvements are achieved through a packed distribution and localized trellis access method, whereby a subset of the full parameter data space is selected, ordered and packed into a new data structure. The new data structure is designed so that the processor can load it into its faster cache memory and then utilize the cached information in a very efficient manner. Specifically, the information is ordered and packed to allow the processor to access the data in substantially sequential order, with a substantially reduced likelihood that the cache will need to be flushed and reloaded (a time consuming and inefficient process).
To understand how the packed distribution and localized trellis access method is able to produce processing speed improvements (10 fold or more) some knowledge of microprocessor caching techniques will be helpful. FIG. 8 illustrates a cached memory architecture of the type generally found in modern day processors. It will be appreciated, however, that FIG. 8 has been simplified to illustrate the caching principle. There are, of course, many different ways to implement caching in various different microprocessor architectures.
The basic concept behind caching is to place the most frequently used program instructions and/or the most frequently used data in the fastest memory available to the processor. In random access memory devices, data access is mediated by a clock. This clock dictates how quickly the information can be read from or written to the memory device under its control. In a typical microprocessor architecture, the microprocessor itself may operate under control of a high speed clock, while the main memory of the computer system will typically operate using a slower clock. This is because it is generally not economically feasible to construct random access main memory circuits that are able to operate at the same clock speed as the microprocessor.
Illustrated in FIG. 8, caches can be implemented at different staged levels, to provide temporary storage for processor instructions and/or data values in memory circuits that are faster than the main memory of the system. Thus, in FIG. 8 main memory 100 is illustrated under the control of clock 102. The so-called level 1 cache or L1 cache is implemented on the microprocessor core, itself, as illustrated at 104. Thus the L1 cache operates under control of clock 106, which also mediates control of the microprocessor. In some microprocessor system designs, additional intermediate stage caches are also included. Illustrated in FIG. 8 is the level 2 cache or L2 cache 108, with its associated clock 110; and level 3 cache or L3 cache 112 with its associated clock 114. It will be recognized, of course, that there are many different cache circuit architectures and thus FIG. 8 is intended to simply introduce the caching concept.
Typically, the memory architecture, illustrated in FIG. 8, acts somewhat as a funnel. Main memory 100 is large, but relatively slow. The L1 cache is comparatively small but operates at very high speed. The intermediate level 2 and level 3 caches are typically smaller than main memory 100, but faster, with the level 2 cache typically being faster than the level 3 cache. These circuits are designed so that information (program instructions and/or data) is automatically loaded from main memory 100 into the successive cache levels, with the hope that processing speed improvements will result. The concept works when the program instructions and data can be loaded as a block and then used for many successive microprocessor core clock cycles without the need to replenish. The caching concept does not work well when the processing algorithm needs program instructions or data that have not been preloaded into the cache, as these will require access to slower memory.
In a typical speech processing application, such as a large vocabulary application, the speech parameters will be stored in a table occupying a portion of the main memory 100. As the speech processing algorithm performs its task, utilizing these parameters, portions of the parameter table will be loaded into the microprocessor's cache memory—as a natural consequence of being accessed by the speech processing algorithm. In conventional speech processing algorithms, however, no attempt is made to optimize what gets loaded into the cache.
According to the present invention, it is possible to optimize what gets loaded into the cache and thereby substantially improve the speed at which speech processing tasks may be performed. Referring to FIG. 9, an example of the packed distribution and localized trellis access technique of the invention will be presented. For purposes of presentation, it has been assumed that the speech application employs parameters in the form of a large Gaussian mixture table. The table contains the Gaussian mixture parameters used to define the states of all Hidden Markov Models for each word or utterance that the system is designed to operate upon. In a typical large vocabulary speech recognition application, the Gaussian mixture table might contain 100,000 different values, expressed as floating points numbers and organized in a predetermined order that is typically based on how the words or utterances are stored in the system's dictionary or lexicon. Although the order in which the Gaussian mixture values are stored is known in advance, it cannot be assumed that this order will correspond to the order with which the mixture values will need to be accessed when the recognition application is used. To understand why this is so, consider how words are stored in an alphabetically arranged dictionary. Although the word order is known, the word order of a sentence using a portion of those words would certainly not be expected to follow the same alphabetical ordering. Rather, the order of words in the sentence would be dictated by grammatical rules and by the semantic requirements of the sentence author.
The present invention selects a subset of the Gaussian mixture table, corresponding to the Gaussian mixture values actually used in the recognition process, and stores that subset in a packed mixture table. In FIG. 9, the entire Gaussian mixture table is illustrated at 120 and the packed mixture table is illustrated at 130. It is the contents of the packed mixture table that will be loaded into the cache 104 as a natural consequence of the speech processing algorithm being used to operate upon the data in the packed mixture table.
Speech data is different from other forms of data, such as financial data, in that there is a sequential order or spoken order to the speech data. This can be illustrated by a directed graph, shown at 122. The graph shows all possible sequences by which one sound unit may follow another sound unit. In this regard, sound units can be individual phones, or they can be larger structures, such as syllables, words, etc. To illustrate the concept of the directed graph, refer to FIG. 10. In FIG. 10 the sound units correspond to individual words that are linked together to form sentences. By traversing generally from left to right, one can construct phrases or sentences such as, “The large roof . . . ” or “The large truck rounded the bend.” In many speech processing applications, such as in a multi-pass recognition application, at some stage the processing algorithm will have knowledge of the temporal order or spoken order of the data being processed. As shown in FIG. 9, knowledge of the order (as depicted diagrammatically using the word/phone graph 122) will establish an access order that the processing algorithm of the invention ascertains at step or module 124. For example, the temporal sequence may require access to Gaussian mixture data values 1 . . . 2 . . . 3 . . . 4. However, these values are (a) not likely to exist sequentially in the Gaussian mixture table and (b) not likely to be stored in contiguous or adjacent memory locations. In FIG. 9, the data element 4 occurs as the first entry in the Gaussian mixture table (reading from top to bottom) whereas the data element 1 is found in the middle of the table.
Rather than utilize the selected Gaussian mixture values directly from mixture table 120, the packed distribution and localized trellis access technique resorts and packs the selected subset of table 120 into the packed mixture table 130. This is performed using the processing step or module 128. The selected subset from table 120 is (a) placed in sequential order that corresponds to the sequential order or spoken order described by graph 122 and (b) packed so that the respective sorted values are adjacent or contiguous in memory.
After performing the resorting and packing operation, the algorithm passes control to the speech processing algorithm that will utilize the data. This is done by passing the address of the packed mixture table 130 to the processing algorithm, so that it will operate upon the data stored in the packed mixture table, rather than upon the data in the Gaussian mixture table 120. In so doing, the microprocessor will load the packed mixture table 130 into its cache 104, where all operations using the cached values will be performed at much higher speed than would be possible if main memory were utilized.
Further expanding on the explanation provided by FIG. 9, refer now to FIG. 11, which gives a specific example where individual words are modeled by HMM Gaussian mixture model parameters and the application is operating upon a spoken utterance, such as a spoken phrase or sentence. In this example, each word may be represented by a plurality of Gaussian mixture parameters, which themselves follow a temporal sequence (e.g., the temporal sequence of the phonemes that make up the word).
In FIG. 11 an exemplary phrase “This is a . . . ” is being processed for recognition. The illustrated spoken order is from left to right. The first spoken word utterance “This” contains three states 1, 2 and 3 having Gaussian mixture parameters stored in the Gaussian mixture table 120 at memory locations 500, 10,000, and 1, respectively. Note that the order of storage within table 120 does not correspond to the temporal sequence of the spoken utterance. This will frequently be the case in speech applications. In contrast, most other data processing applications, such as financial applications will frequently be able to access stored data in the order in which it was stored.
The operation of the re-sort and pack step or module 128 (FIG. 9) is shown by the sorting and storing lines 140 in FIG. 11. The Gaussian mixture value stored at memory location 500 in table 120 is copied to memory location 1 in the packed mixture table 130. Next, the mixture values stored at location 10,000 is copied and stored at location 2 within the packed mixture table 130. Finally, the value stored at memory location 1 in table 120 is copied to memory location 3 in table 130. Thereafter, memory location 4 in table 130 would be populated with the data value that corresponds to the first state of the next occurring spoken utterance (the word “is”). Thus note that the table 130 contains data that is ordered based on the established access order dictated by the spoken order of the input utterance, and that the data values are packed in contiguous memory locations.
For a further understanding of the presently preferred processing method, refer to FIG. 12. FIG. 12 illustrates a method for processing speech data utilizing high speed cache memory having an associated cache mechanism for transfer of data from system memory into cache memory. The order of the steps illustrated in FIG. 12 may be varied from what is shown there, without departing from the spirit of the invention as more fully set forth in the appended claims. At step 200 a main table of speech parameters is provided. These speech parameters are preferably stored in the main system memory (main memory 100, FIG. 8). In addition, a list is provided at step 202 to establish a processing order for at least a portion of the speech parameter data stored in the main table. The parameters stored in the main table may be speech parameters, such as Gaussian mixture parameters associated with corresponding Hidden Markov Models, and the list provided at 202 may be in the form of a set of sequential data corresponding to a spoken utterance, or some other sequence that has a temporal structure.
At step 204, data items are selected from the main table based on the list order and these are copied into a sub-table. As illustrated by constraining steps 206 and 208, the sub-table is processed so that entries are stored in contiguous memory locations. In the presently preferred embodiment, the sub-table may be implemented in main system memory. In addition, the sub-table is processed so that entries are sorted according to the processing order established by the list in step 202. It will be appreciated that constraining steps 206 and 208 can be processed in either order, or concurrently. In the presently preferred embodiment the sub-table is constructed by sequentially adding entries to the sub-table in successively contiguous memory locations, with the order of the entries being established by selecting them from the main table in the order established by the list. Of course, alternate embodiments can be envisioned where the sub-table is initially constructed in a non-contiguous fashion and then later compacted, or where the sub-table is initially constructed with one sort order and thereafter re-sorted according to the list.
After copying the entries to the sub-table in step 204, the applicable speech processing algorithm is then used to operate upon the sub-table at step 210. By operating upon the sub-table, the sub-table is transferred into high speed cache memory by utilizing the cache mechanism associated with the cache memory. In this regard, most modern day microprocessors will automatically transfer a given block of information into high speed cache memory, so that that block of data can be processed more rapidly. Of course, the transfer into high speed cache memory can also be effected by an explicit processor command, if desired.
The packed distribution and localized trellis access method illustrated in FIGS. 8-11
and described above can be advantageously used in a variety of different speech processing applications. Examples of such applications include:
- local distance computation and trellis expansion for algorithms of the Viterbi and Baum-Welch type. These computations are central to many large vocabulary continuous speech recognition training and recognition applications;
- Viterbi beam search algorithms for real-time recognition;
- constrained search on word/phone graphs or focused language models;
- re-scoring of word lattices for acoustic model adaptation;
- maximum mutual information estimation (MMIE) acoustic model training;
- expectation maximization-based maximum likelihood acoustic model training;
- multi-pass recognition processes.
In applications such as those listed above, the packed distribution and localized trellis access method provides a speed improvement of at least one order of magnitude over conventional methods. The exact speed improvement factor depends, in part, upon the speed benefit that the high speed cache memory produces over system memory. Thus, processors having faster high-speed cache performance will show even greater speed improvement when the technique is utilized. In this regard, it will be appreciated that the technique exploits the speed benefit of the high-speed cache by (a) localizing the memory access based upon the order that the expansion algorithm or other speech processing algorithm explores the trellis and (b) sorting the memory representation of the Gaussian parameters (or other speech parameters) such that the memory is accessed in increasing order.
In general, the advantages of the technique can be enjoyed in applications where the system has some a priori knowledge of which speech parameters will be needed, and that those parameters be of sufficiently small size as to fit in the high-speed cache memory. For a typical very large vocabulary continuous speech recognition application, these conditions are met during training and during passes of recognition that occur after a first pass (e.g., adaptation or re-scoring passes). These conditions are also met for dialog systems where each state is associated with a particular vocabulary, and for text-prompted speaker recognition systems, where each prompted text evokes a particular set of speech parameters. Finally, the algorithm described here can be combined with the other algorithms described earlier in this document to further improve memory access by decreasing the bandwidth used between the main CPU and the system memory.
The foregoing discloses and describes merely exemplary embodiments of the present invention. One skilled in the art will readily recognize from such discussion, and from accompanying drawings and claims, that various changes, modifications, and variations can be made therein without departing from the spirit and scope of the present invention.