State-Dependent Speaker Clustering for Speaker Adaptation

The invention relates to speaker adaptation in speech recognition systems.

Speech recognition systems are based on using the correlation between a spoken word and the characterization of that word in the acoustic space. These characterizations are typically obtained from training data. The training data could be derived from a number of speakers, which would result in a speaker independent system, or it could be derived from a single speaker, the same speaker who would be using the system. In the latter case, the system is said to be speaker-dependent. Typically, speaker-independent recognition systems recognize speech less accurately than speaker-dependent systems (by about 35% for continuous speech). This is because speaker-independent systems try to capture the average statistics across several speakers, and may not characterize any single speaker particularly well. However, speaker-dependent systems have the disadvantage that they require the test speaker to provide a significant amount of training data.

It is possible to improve upon speaker-independent performance to some extent, without requiring the test speaker to train-up a speaker-dependent system. As the speaker-independent system is trained on a number of speakers, the statistics reflect the variations among all these speakers. To some extent, the statistics can be made more specific by making a gender-dependent system, which corresponds to dividing the space of test speakers into two sets, and training the system to model the characteristics of each set separately. This idea could also be extended further by making up several speaker clusters, rather than just two. However, as the number of speakers in the training data is limited, improvement using such means is usually marginal.

An alternative to these methods for the purpose of sharpening the characteristics of the system to better match a test speaker, is speaker adaptation. With speaker adaptation, a few sentences from the test speaker are used to tailor the system to the speaker. So far, speaker adaptation techniques have taken on two forms: in a first scheme, the test speaker's data is transformed to better match the model statistics (see, e.g., J. R. Bellegarda, P. V. de Souza, A. Nadas, D. Nahamoo, M. A. Picheny and L. R. Bahl, "The Metamorphic Algorithm: A Speaker Mapping Approach to Data Augmentation", IEEE Transactions on Speech and Audio Processing, vol. 2, no. 3, pp 413-420, July 1994 ("Bellegarda")); in the second scheme, the model is tailored to better fit the test speaker's characteristics (see e.g., C. J. Legetter and P. C. Woodland, "Speaker Adaptation of Continuous density HMM's using Multivariate Linear Regression", ICSLP 194, vol. 2, pp 451-454 ("Legetter"); George Zavaliagkos, "Bayesian Adaptation Techniques for Speaker Adaptation", Ph.D thesis, Northeastern University, 1995 ("Zavaliagkos"); V. Digalakis and L. Neumeyer, "Speaker Adaptation using Combined Transformation and Bayesian techniques", Proceedings of the Intl. Conf. Acoustics, Speech and Signal Processing, May, 1995).

There are several techniques that have evolved within the framework of the second scheme. For instance, in Legetter and Digilakis the new model parameters are assumed to be related to the speaker-independent parameters via a linear transformation, and the transform is computed to maximize the likelihood of the test speaker's adaptation data. Alternatively, it is possible to accumulate statistics from the test speaker's adaptation data and re-estimate the system parameters by smoothing these statistics back to the speaker-independent parameters, which is equivalent to computing a maximum-a-posteriori estimate of the system parameters (see Zavaliagkos and J. L. Gauvain and C. H. Lee, "Maximum-a-posteriori Estimation for Multivariate Gaussian Observations of Markov Chains", IEEE Transactions on Speech and Audio Processing, vol. 2, no. 2, pp 291-298, April 1994 (here, the inter-speaker variability in the training data is characterized by a prior distribution on the parameters of the system, and this prior distribution is used in conjunction with the collected statistics to re-estimate the parameters)).

In this invention, we present an adaptation technique based on speaker clustering. Typically, the drawback of speaker clustering is that one needs a very large number of training speakers in order to ensure that at least one cluster of training speakers matches the test speaker, i.e., the training speakers in this cluster are acoustically very similar to the test speaker. This drawback makes the straightforward technique of speaker clustering impractical. However, this invention presents a clustering technique that can be used with relatively very few training speakers.

This invention is prompted by the observation that a particular training speaker may be close to the test speaker in a certain region of the acoustic space, but far away in other regions, i.e., two speakers may pronounce the phone AA in the same way, but the phone R in an entirely different way. In such situations, it is obviously incorrect to conclude that the two speakers are close together in general. In order to solve this problem, this invention proposes the idea that the set of training speakers deemed to be close to a test speaker should be determined independently for different sounds. Hence, we would have independent sets of training speakers that characterize different sounds of the test speaker, and the acoustic prototype characterizing the sounds of the test speaker would be made from data obtained from these independent training speaker sets. This invention also proposes the extension of this idea to model other aspects of the system statistics, for instance the modelling of the duration of certain sounds.

An objective of the invention is to take a few sentences of the test speaker's data and use this data to obtain a subset of training speakers who are "close" to the test speaker, with the list of speakers being possibly different for different regions of the acoustic space. The models corresponding to a region of the acoustic space are then remade (i.e., means and variances of the Gaussians are re-estimated) based only on data from the subset of training speakers for that acoustic subspace, in order to produce a model that is a sharper fit to the test speaker. Further, the data from each training speaker can be transformed (using linear (see Legetter) or non-linear (see Bellagarda) technique to bring it even closer to the test speaker, for the particular sound. The invention first calculates the acoustic characterization of the training speakers for all regions of the acoustic space (possibly corresponding to phone classes) from the training speakers' data. Subsequently, a few sentences of the test speaker are used to acoustically characterize the same regions of the acoustic space. A distance score is now computed between the test speaker's models for an acoustic subspace, and the training speakers' model for the same acoustic subspace, for all training speakers. The training speakers are now ranked on the basis of this score. Note that the ranking could be different for different phone classes. Subsequently, the acoustic models for a particular phone class are remade using only data from those training speakers who are the closest to the test speaker for that phone class.

The invention also extends this idea to apply to other aspects of the system models. For instance, because many speakers vary in the length of time over which they pronounce a sound, the speakers could also be characterized by the duration of the phone classes, and the similarity between a test speaker and a training speaker could be determined on the basis of the similarity in the length distribution for the phone class. Hence, the duration model for a particular phone class is remade based on the duration models of only those training speakers who have been selected as being "close" to the test speaker for that phone class.

One embodiment of the invention in the above form will iterate twice in using the training speakers' data; in the first step, the shortlist of training speakers who are deemed to be close to the test speaker is generated, and in the second step, the training data is re-used from the speakers in the shortlist, in order to remake various aspects of the model to better fit the test speaker.

For the case where the acoustic space is modeled as a mixture of Gaussians, the means and variances of the Gaussians are re-estimated using the data from the short list of training speakers. Note that if training speaker 1 has been selected as being close to the test speaker in the acoustic space of phone AA, then only the subset of the training speaker's data that is aligned to the phone AA will be used in the re-estimation process. Further, as mentioned earlier, it is possible to compute a linear (Legetter) or non-linear (Jerome) transform that maps to the training speaker's data in the selected acoustic base to be even closer to the test speaker space.

As far as other model aspects are concerned, as for example the duration of phone classes, the same process as above is followed to select the subset of the training data that is used to re-estimate the model parameters.

It is also possible to avoid having to reuse the training data by storing the models of the training speakers for various phone classes, and simply assembling these training speaker models to make up the new model for the test speaker.

FIG. 1 shows tree diagrams corresponding to a plurality of different contexts of the phone AA.

FIG. 2 is a flow diagram of a method for generating training speaker models.

FIG. 3 is a flow diagram of a method for generating test speaker models.

FIG. 4 is a flow diagram of a method for computing the distance between the training speaker and test speaker models.

FIG. 5 is a flow diagram depicting a method for modifying the prototypes of a speaker independent system using speaker shortlists available for selected nodes.

FIG. 6 is a block diagram of a system in accordance with the invention.

In accordance with the invention, the acoustic space is divided into several regions, and the characteristics of each training speaker are modelled for each region. In order to be able to find the distance between the test speaker and each training speaker, models are generated for the test speaker for the same regions. As far as the training speakers are concerned, it is preferable that there be enough data to generate separate models for several small regions of the acoustic space, such as a phone in a particular context (henceforth, these regions will be referred to as leaves). Note that these training speaker models differ from the speaker-independent models that are finally used in the decoder, which typically model the leaves by much more detailed models. Because only a few sentences of test speaker data have been obtained, it may not be possible to generate test speaker models for all leaves; for instance, it may be possible to characterize the test speaker only at the level of phones (regardless of context), or groups of phones. Consequently, it is useful to have a hierarchical scheme for constructing the training speaker models, depending on how the test speaker data can be characterized.

An example that demonstrates this hierarchical scheme is depicted in FIG. 1. The phone AA could be characterized, for instance, by 50 different contexts (AA.sub.-- 1 through AA.sub.-- 50), and a model could therefore be made for each training speaker for each of these contexts. These contexts correspond to the terminal leaves of the tree in FIG. 1. These models can also be merged to form models for groups of phones-in-context, as shown in FIG. 1. The rules governing the merging of these models can be derived by a bottom-up clustering procedure, as in FIG. 1, or by alternative means. Hence, in FIG. 1, node 51 represents a region of the acoustic space that contains instances of the phone AA in two different contexts, AA.sub.-- 1 and AA.sub.-- 2, and a model could be constructed to represent node 51 by combining the models for nodes 1 and 2. Similarly, the acoustic space representing nodes 51 and 3 can be combined to form node 52.

A method for constructing the training speakers' models is depicted in FIG. 2. For each speaker (step 202), the data aligning to each leaf l (step 204) is collected, and a model is constructed for the leaf (step 206). One possible way to model the acoustic feature vectors is to use a mixture of Gaussians. If length distributions are used, one possible modeling technique is by approximating the length distribution by a poisson distribution.

Let S=›s.sub.1, . . . ,s.sub.S ! represent the set of all training speakers in the training data. Further, let G.sub.l.sup.i represent the model for leaf l of training speaker i (one possible model is a single Gaussian).

The manner in which the test speaker's models are constructed is depicted in FIG. 3. The test speakers adaptation data, which consists of acoustic feature vectors for the test speaker, is first aligned at step 300 using the viterbi alignment algorithm, which enables every vector to be tagged by a leaf, and the alignments are stored.

The counts of the number of feature vectors corresponding to the different leaves of a tree such as the one in FIG. 1 is obtained (step 302). If the count at any leaf is insufficient to allow reliable construction of a model for that leaf, where reliability can be based, for instance, on a threshold count, then the hierarchy shown in FIG. 1 is used to determine the node upstream from the inadequate leaf at which there is sufficient test speaker data to reliably estimate a model (step 304). Hence, for inadequate leaves the test speaker is characterized by a model comprising these reliable nodes, rather than the leaves. The reliable nodes could represent phones, or groups of phones. Next, the training speaker models must also be constructed for these same (reliable) nodes, and this is done by assembling the models for all the leaves which form the node, according to the hierarchy of FIG. 1.

Let ›n.sub.1 . . . , n.sub.N ! represent the nodes of the tree for which test speaker models are made, and let L.sub.i =›l.sub.l.sup.i, . . . , l.sub.kl.sup.i ! represent the leaves that are shared to form node n.sub.i, according to the hierarchy of FIG. 1.

It now remains to compute the distance between the models of the test speaker and each training speaker, for the corresponding nodes. This can be done in one of several ways. For example, the test speaker model for a node could simply be taken to be all the feature vectors that align to the node, and each training speaker's model for the node could be made up as a mixture of Gaussians corresponding to all the leaves that are shared at that node. Hence, a measure of the closeness between the test speaker and a training speaker could be simply the likelihood of the test speaker's data at that node, computed using the training speaker's mixture of Gaussians for that node. Each training speaker is then ranked according to this distance, and a shortlist of training speakers is produced representing the training speakers, that are closest to the test speaker for that node. Let S.sub.i denote the shortlist of speakers corresponding to node n.sub.i.

This shortlist generation process is depicted in FIG. 4. Here, Y.sub.t.sup.i represents the acoustic feature vector of the t.sup.th frame of test speaker data (for notational convenience, all sentences of the test speaker are assumed to have been concatenated into one sentence), and the superscript i indicates that this frame has been aligned to leaf i. The test speaker data is viterbi aligned against the script (this could be the correct script corresponding to supervised adaption, or the script obtained by a first pass decoding, corresponding to unsupervised adaptation.) and each acoustic feature vector is assigned a tag that is a leaf id.

Now, as mentioned earlier, there is insufficient test speaker data to enable a modeling of each leaf, hence, groups of leaves are combined in accordance with FIG. 1 to form a model at a node that includes these groups of leaves, which correspond to nodes n.sub.i in the tree of FIG. 1. Hence, each leaf id i in the test speakers alignment can be replaced by the node n.sub.i which represents the node in the tree of FIG. 1 that includes leaf i, and for which there is sufficient data to make up a model for the test speaker. This is done in block 404.

Hence, let y.sub.t.sup.ni represent the vector at time t, that has been tagged with node n.sub.i (which refers to a node in the tree of FIG. 1). The term Dist(Y.sub.t.sup.ni /Model.sub.ni.sup.j) represents the distance between the acoustic feature vector Y.sub.t.sup.ni and the model of training speaker j for node n.sub.i. One possibility is to make Model.sub.ni.sup.j =G.sub.ni.sup.j, i.e., use a single Gaussian. In this case, the distance may be computed simply as the likelihood, or as the euclidean distance between the feature vector Y.sub.t.sup.ni and the mean of the Gaussian.

Thus, referring to FIG. 4, the process is as follows. At step 400, D.sub.ni.sup.j is initialized to 0 for each j.epsilon.S. Next, for all t, and for a speaker j, compute: a) the distance d.sub.j between acoustic vector Y.sub.t.sup.ni, and the model (Model.sub.ni.sup.j) of training speaker j for node n.sub.i ; and b) add these dj together to form the distance D.sub.ni.sup.j between the test speaker and training speaker j for node n.sub.i (steps 402, 404 and 406). Repeat this for all training speakers j.epsilon.S. Next, for all nodes n.sub.i, the training speakers are ranked on the basis of the values of D.sub.ni.sup.j, and a shortlist of training speakers for node n.sub.i is compiled, comprising the training speakers with the lowest values of D.sub.ni.sup.j (steps 408 and 410).

Once a shortlist of speakers has been selected for each n.sub.i, the data from this shortlist of speakers can be directly used to re-estimate the parameters of the Gaussians that characterize the acoustic region corresponding to node ni. Further, instead of using the training speakers data directly, it is also possible to compute a transformation for every training speaker in the shortlist that when applied to the training speakers data, will bring the training speakers data (in the acoustic subspace corresponding to node n.sub.i) even closer to the test speakers space.

Either a linear or non-linear transformation can be used for this purpose, as in Legetter or Bellagarda. For example, a linear transformation can be computed to maximize the likelihood of the test speaker data that has aligned to node ni given the j.sup.th training speakers model for node ni. The transformation that maps the acoustic subspace corresponding to node ni for the j.sup.th training speaker closer to the subspace corresponding to node n.sup.i of the test speaker is denoted T.sub.ni.sup.j and computed in block 412.

The only remaining issue is how to modify the prototypes of the speaker-independent system using the speaker shortlists that are available for selected nodes in the hierarchy of FIG. 1. This process is shown in FIG. 5. As mentioned earlier, the test speaker is characterized by a model at several nodes, and each of these nodes corresponds to grouping together a number of leaves and/or nodes. The speaker-independent system typically models the acoustics at each leaf by a mixture of Gaussians (the length distribution can be modelled by imposing a topology on the hidden Markov model for the leaf, e.g., the model for a leaf should be at least seven feature vectors long). Let N.sub.i =›N.sub.l.sup.i, . . . , N.sub.ni.sup.i ! represent the Gaussians of the training speakers that correspond to leaf i.

In FIG. 4, it was shown how to obtain a shortlist of training speakers who are close to the test speaker in the acoustic space corresponding to a particular node. This shortlist of speakers is used to modify the prototypes of the speaker-independent system corresponding to that node, as shown in FIG. 5. In this figure, M.sub.j.sup.i represents a vector V.sub.j.sup.i represents a matrix, and C.sub.j.sup.i represents a scalar, X.sub.s,t.sup.i represents the t.sup.th frame of training speaker s, aligned to leaf i. Prob(X.sub.s,t.sup.i /N.sub.j.sup.i) indicates the likelihood of the feature vector, given the Gaussian N.sub.j.sup.i. Y.sub.s,t.sup.i represents the feature vector obtained after applying the transformation T.sub.n.sbsb.i.sup.s to X.sub.s,t and Y.sub.s,t.sup.i2 meant to indicate that each element of the vectorY.sub.s,t.sup.i has been squared. Note that it is possible not to use the transformation, in which case Y.sub.s,t.sup.i =X.sub.s,t.sup.i. Once the prototypes have been remade, the remainder of the test speakers data are decoded with the updated prototypes.

FIG. 5 will now be described in detail. In order to estimate the means and co-variances of the Gaussians, it is first necessary to compute the moments of the data. By moments, we mean the following: p.sub.j.sup.i, the posterior probability of a feature vector being generated by the j.sup.th Gaussian modeling leaf i. (Note that ##EQU1## The first moment given by equation ##EQU2## where Y.sub.s,t is the transformed feature vector at time t for the s.sup.th training speaker, and the second moment ##EQU3## For the case where the co-variance matrix is assumed to be diagonal, V.sub.j.sup.i can be replaced by ##EQU4## where Y.sub.s,t.sup.2 denotes a vector made up by squaring every element in the vector Y.sub.s,t. For each node n.sub.i, the process uses the data from the shortlist of training speakers made up for node n.sub.i to re-estimate the Gaussian parameters of the speaker-independent model corresponding to node n.sub.i. (Blocks 501-504). For our case, as we already have the leaf-level alignments of the training data, the (p.sub.j.sup.i) can be computed as ##EQU5## where i is the leaf to which the current feature vector has been aligned, and the index j denotes the j.sup.th prototype used to model the data at that leaf. In order to re-estimate the Gaussian parameters, it is necessary to compute the moments of the data. The training speakers data can be directly used for this purpose, or the transformed training speakers data can be used for this purpose. As shown in FIG. 4, it is possible to compute a transformation (linear or non-linear) that maps the training speakers space closer to the test speaker in the acoustic subspace corresponding to node n.sub.i. Let us denote the feature vector after applying the transformation as y.sub.s,t.sup.i (s denotes the s.sup.th training speaker, t denotes the t.sup.th feature vector, i denotes the leaf to which the feature vector was aligned and the transform T.sub.n.sbsb.i.sup.s was applied to x.sub.s,t.sup.i to transform it to y.sub.s,t.sup.i (Block 505).

The moments of the data are computed using the transformed feature vector Y.sub.s,t.sup.i (Block 508) and the Gaussian parameters re-estimated from the moments (Block 511).

FIG. 6 is a system diagram of system implementing the method of the invention. A system in accordance with the invention includes a feature store 602, viterbi alignment processor 604, speaker training data model 606, feature vector store 608, viterbi alignment processor 610, best speaker model 611, training speaker select processor 612, training data transformer 614, Gaussian parameter re-estimation processor 616 and decoder 618. The feature vector and script store 602 stores feature vectors of training speakers along with the accompanying script used to produce those feature vectors. The viterbi alignment processor 604 uses the script and corresponding feature vectors to assign classes to each feature vector using hidden Markov models. The training speaker data model collects feature vectors that belong to particular leaves for a particular training speaker and computes a model, (e.g., a Gaussian) based on the data at each leaf. The feature vector store 608 collects and stores feature vectors and the script of the test speaker. The script can be the correct script, or a script produced by prior decoding. The viterbi alignment processor 610 computes the viterbi alignment and 611 the test speaker model (i.e., nodes for which the test model can be made based only on the test data available). Training speaker select processor 612 is used to select training speakers that are closest to the test speaker for a selected node in the tree. Training data transformer 614 applies a transform on the selected training speakers to better map the training data to the test speaker's acoustic space. Gaussian parameter re-estimation processor 616 estimates the first and second moment and posterior probability of a transformed feature vector belonging to a Gaussian. From these, it is possible to compute means and variances. Decoder 618 decodes the test speakers data using the new Gaussians estimated in block 616.

The features of FIG. 16 can be implemented in a general purpose computer.