In this paper, we describe basis-based speaker adaptation techniques using the matrix representation of training models. Bases are obtained from training models by decomposition techniques for matrix-variate objects: two-dimensional principal component analysis (2DPCA) and generalized low rank approximations of matrices (GLRAM). The motivation for using matrix representation is that the sample covariance matrix of training models can be more accurately computed and the speaker weight becomes a matrix. Speaker adaptation equations are derived in the maximum-likelihood (ML) framework, and the adaptation equations can be solved using the maximum-likelihood linear regression technique. Additionally, novel applications of probabilistic 2DPCA and GLRAM to speaker adaptation are presented. From the probabilistic 2DPCA/GLRAM of training models, speaker adaptation equations are formulated in the maximum a posteriori (MAP) framework. The adaptation equations can be solved using the MAP linear regression technique. In the isolated-word experiments, the matrix representation-based methods in the ML and MAP frameworks outperformed maximum-likelihood linear regression adaptation, MAP adaptation, eigenvoice, and probabilistic PCA-based model for adaptation data longer than 20 s. Furthermore, the adaptation methods using probabilistic 2DPCA/GLRAM showed additional performance improvement over the adaptation methods using 2DPCA/GLRAM for small amounts of adaptation data.

We present a tensor analysis of acoustic models comprising various speakers in multiple noise conditions, and its application to the new speaker and environment adaptation for speech recognition. The bases used in adaptation are constructed by decomposing the training models in the state, feature dimension, speaker, and noise spaces using multilinear singular value decomposition. The isolated-word recognition experiment demonstrated the effectiveness of the proposed method, showing better performance than eigenvoice in the babble and factory floor noises for the adaptation data longer than approximately 20 s.

This paper presents a novel speaker adaptation method based on the multilinear analysis of training speakers using Tucker decomposition. A Tucker decomposition of training models can decouple the dataset into the subspaces of state, dimension of the mean vector, and speaker. Using the bases of the state subspace, we derive a speaker adaptation formula where the matrix of basis vectors is weighted in row and column spaces; the proposed method can include the eigenvoice technique as a subset. The results from the isolated-word recognition task showed that the Tucker decomposition-based method outperformed both eigenvoice and MLLR for the adaptation data whose lengths are 15 seconds or longer. Furthermore, the method can easily be extended to multi-factor problems, thus enabling the adaptation of multiple factors such as speaker and noise environment.

This letter describes a speaker adaptation method based on the two-dimensional PCA of training models. In the method, state and dimension of mean vectors are differentiated, and the covariance matrix is computed dimension-wisely. As a result, the speaker weight can contain different weighting for each dimension of mean vectors. In the isolated-word recognition experiments, the proposed method performed better than both eigenvoice and MLLR, for adaptation data longer than about 15 seconds, due to its more elaborate modeling. The method can also be applied to other PCA-based modeling methods where each training model can be represented as a matrix.

In this paper, a novel method for speaker adaptation using bilinear model is proposed. Bilinear model can express both characteristics of speakers (style) and phonemes across speakers (content) independently in a training database. The mapping from each speaker and phoneme space to observation space is carried out using bilinear mapping matrix which is independent of speaker and phoneme space. We apply the bilinear model to speaker adaption. Using adaptation data from a new speaker, speaker-adapted model is built by estimating the style(speaker)-specific matrix. Experimental results showed that the proposed method outperformed eigenvoice and MLLR. In vocabulary-independent isolated word recognition for speaker adaptation, bilinear model reduced word error rate by about 38% and about 10% compared to eigenvoice and MLLR respectively using 50 words for adaptation.

The segmentation of deformable objects from three-dimensional (3-D) images is an important and challenging problem, especially in the context of medical imagery. We present a new segmentation algorithm based on matching probability distributions of photometric variables that incorporates learned shape and appearance models for the objects of interest. The main innovation over similar approaches is that there is no need to compute a pixelwise correspondence between the model and the image. This allows for a fast, principled algorithm. We present promising results on difficult imagery for 3-D computed tomography images of the male pelvis for the purpose of image-guided radiotherapy of the prostate.

We present a computer vision tool to improve the clinical outcome of patients undergoing radiation therapy for prostate cancer by improving irradiation technique. While intensity modulated radiotherapy (IMRT) allows one to irradiate a specific region in the body with high accuracy, it is still difficult to know exactly where to aim the radiation beam on every day of the 30~40 treatments that are necessary. This paper presents a geometric model-based technique to accurately segment the prostate and other surrounding structures in a daily serial CT image, compensating for daily motion and shape variation. We first acquire a collection of serial CT scans of patients undergoing external beam radiotherapy, and manual segmentation of the prostate and other nearby structures by radiation oncologists. Then we train shape and local appearance models for the structures of interest. When new images are available, an iterative algorithm is applied to locate the prostate and surrounding structures automatically. Our experimental results show that excellent matches can be given to the prostate and surrounding structure. Convergence is declared after 10 iterations. For 256 x 256 images, the mean distance between the hand-segmented contour and the automatically estimated contour is about 1.5 pixels (2.44 mm), with variance about 0.6 pixel (1.24 mm).