Fast Likelihood Computation in Speech Recognition using Matrices

Acoustic modeling using mixtures of multivariate Gaussians is the prevalent approach for many speech processing problems. Computing likelihoods against a large set of Gaussians is required as a part of many speech processing systems and it is the computationally dominant phase for Large Vocabulary Continuous Speech Recognition (LVCSR) systems. We express the likelihood computation as a multiplication of matrices representing augmented feature vectors and Gaussian parameters. The computational gain of this approach over traditional methods is by exploiting the structure of these matrices and efficient implementation of their multiplication. In particular, we explore direct low-rank approximation of the Gaussian parameter matrix and indirect derivation of low-rank factors of the Gaussian parameter matrix by optimum approximation of the likelihood matrix. We show that both the methods lead to similar speedups but the latter leads to far lesser impact on the recognition accuracy. Experiments on 1,138 work vocabulary RM1 task and 6,224 word vocabulary TIMIT task using Sphinx 3.7 system show that, for a typical case the matrix multiplication based approach leads to overall speedup of 46 % on RM1 task and 115 % for TIMIT task. Our low-rank approximation methods provide a way for trading off recognition accuracy for a further increase in computational performance extending overall speedups up to 61 % for RM1 and 119 % for TIMIT for an increase of word error rate (WER) from 3.2 to 3.5 % for RM1 and for no increase in WER for TIMIT. We also express pairwise Euclidean distance computation phase in Dynamic Time Warping (DTW) in terms of matrix multiplication leading to saving of approximately of computational operations. In our experiments using efficient implementation of matrix multiplication, this leads to a speedup of 5.6 in computing the pairwise Euclidean distances and overall speedup up to 3.25 for DTW.

Diversity-Rate Trade-off for Fading Channels with Power Control

We consider a complex, additive, white Gaussian noise channel with flat fading. We study its diversity order vs transmission rate for some known power allocation schemes. The capacity region is divided into three regions. For one power allocation scheme, the diversity order is exponential throughout the capacity region. For selective channel inversion (SCI) scheme, the diversity order is exponential in low and high rate region but polynomial in mid rate region. For fast fading case we also provide a new upper bound on block error probability and a power allocation scheme that minimizes it. The diversity order behaviour of this scheme is same as for SCI but provides lower BER than the other policies.

Nodal length fluctuations for arithmetic random waves

Using the spectral multiplicities of the standard torus, we endow the Laplace eigenspaces with Gaussian probability measures. This induces a notion of random Gaussian Laplace eigenfunctions on the torus (''arithmetic random waves''). We study the distribution of the nodal length of random eigenfunctions for large eigenvalues, and our primary result is that the asymptotics for the variance is nonuniversal. Our result is intimately related to the arithmetic of lattice points lying on a circle with radius corresponding to the energy.

Semi-supervised SVMs for classification with unknown class proportions and a small labeled dataset

In the design of practical web page classification systems one often encounters a situation in which the labeled training set is created by choosing some examples from each class; but, the class proportions in this set are not the same as those in the test distribution to which the classifier will be actually applied. The problem is made worse when the amount of training data is also small. In this paper we explore and adapt binary SVM methods that make use of unlabeled data from the test distribution, viz., Transductive SVMs (TSVMs) and expectation regularization/constraint (ER/EC) methods to deal with this situation. We empirically show that when the labeled training data is small, TSVM designed using the class ratio tuned by minimizing the loss on the labeled set yields the best performance; its performance is good even when the deviation between the class ratios of the labeled training set and the test set is quite large. When the labeled training data is sufficiently large, an unsupervised Gaussian mixture model can be used to get a very good estimate of the class ratio in the test set; also, when this estimate is used, both TSVM and EC/ER give their best possible performance, with TSVM coming out superior. The ideas in the paper can be easily extended to multi-class SVMs and MaxEnt models.

Fast acoustic likelihood computation using low-rank matrix approximation

Acoustic modeling using mixtures of multivariate Gaussians is the prevalent approach for many speech processing problems. Computing likelihoods against a large set of Gaussians is required as a part of many speech processing systems and it is the computationally dominant phase for LVCSR systems. We express the likelihood computation as a multiplication of matrices representing augmented feature vectors and Gaussian parameters. The computational gain of this approach over traditional methods is by exploiting the structure of these matrices and efficient implementation of their multiplication.In particular, we explore direct low-rank approximation of the Gaussian parameter matrix and indirect derivation of low-rank factors of the Gaussian parameter matrix by optimum approximation of the likelihood matrix. We show that both the methods lead to similar speedups but the latter leads to far lesser impact on the recognition accuracy. Experiments on a 1138 word vocabulary RM1 task using Sphinx 3.7 system show that, for a typical case the matrix multiplication approach leads to overall speedup of 46%. Both the low-rank approximation methods increase the speedup to around 60%, with the former method increasing the word error rate (WER) from 3.2% to 6.6%, while the latter increases the WER from 3.2% to 3.5%.

Capacity region of K-user discrete memoryless interference channels with a mixed strong-very strong interference

The capacity region of the 3-user Gaussian Interference Channel (GIC) with mixed strong-very strong interference was established in [1]. The mixed strong-very strong interference conditions considered in [1] correspond to the case where, at each receiver, one of the interfering signals is strong and the other is very strong. In this paper, we derive the capacity region of K-user (K ≥ 3) Discrete Memoryless Interference Channels (DMICs) with a mixed strong-very strong interference. This corresponds to the case where, at each receiver one of the interfering signals is strong and the other (K - 2) interfering signals are very strong. This includes, as a special case, the 3-user DMIC with mixed strong-very strong interference. The proof is specialized to the 3-user GIC case and hence an alternative derivation for the capacity region of the 3-user GIC with mixed strong-very strong interference is provided.

Channel estimation using minimum bit error rate framework for BPSK signals

We consider the design of a linear equalizer with a finite number of coefficients in the context of a classical linear intersymbol-interference channel with additive Gaussian noise for channel estimation. Previous literature has shown that Minimum Bit Error Rate(MBER) based detection has outperformed Minimum Mean Squared Error (MMSE) based detection. We pose the channel estimation problem as a detection problem and propose a novel algorithm to estimate the channel based on the MBER framework for BPSK signals. It is shown that the proposed algorithm reduces BER compared to an MMSE based channel estimation when used in MMSE or MBER detection.

A hybrid RTS-BP algorithm for improved detection of large-MIMO M-QAM signals

Low-complexity near-optimal detection of large-MIMO signals has attracted recent research. Recently, we proposed a local neighborhood search algorithm, namely reactive tabu search (RTS) algorithm, as well as a factor-graph based belief propagation (BP) algorithm for low-complexity large-MIMO detection. The motivation for the present work arises from the following two observations on the above two algorithms: i) Although RTS achieved close to optimal performance for 4-QAM in large dimensions, significant performance improvement was still possible for higher-order QAM (e.g., 16-, 64-QAM). ii) BP also achieved near-optimal performance for large dimensions, but only for {±1} alphabet. In this paper, we improve the large-MIMO detection performance of higher-order QAM signals by using a hybrid algorithm that employs RTS and BP. In particular, motivated by the observation that when a detection error occurs at the RTS output, the least significant bits (LSB) of the symbols are mostly in error, we propose to first reconstruct and cancel the interference due to bits other than LSBs at the RTS output and feed the interference cancelled received signal to the BP algorithm to improve the reliability of the LSBs. The output of the BP is then fed back to RTS for the next iteration. Simulation results show that the proposed algorithm performs better than the RTS algorithm, and semi-definite relaxation (SDR) and Gaussian tree approximation (GTA) algorithms.

A Huber-loss-driven clustering technique and its application to robust cell detection in confocal microscopy images

We address the problem of detecting cells in biological images. The problem is important in many automated image analysis applications. We identify the problem as one of clustering and formulate it within the framework of robust estimation using loss functions. We show how suitable loss functions may be chosen based on a priori knowledge of the noise distribution. Specifically, in the context of biological images, since the measurement noise is not Gaussian, quadratic loss functions yield suboptimal results. We show that by incorporating the Huber loss function, cells can be detected robustly and accurately. To initialize the algorithm, we also propose a seed selection approach. Simulation results show that Huber loss exhibits better performance compared with some standard loss functions. We also provide experimental results on confocal images of yeast cells. The proposed technique exhibits good detection performance even when the signal-to-noise ratio is low.

Frequency detection from multiplexed compressed sensing of noisy signals

Signal acquisition under a compressed sensing scheme offers the possibility of acquisition and reconstruction of signals sparse on some basis incoherent with measurement kernel with sub-Nyquist number of measurements. In particular when the sole objective of the acquisition is the detection of the frequency of a signal rather than exact reconstruction, then an undersampling framework like CS is able to perform the task. In this paper we explore the possibility of acquisition and detection of frequency of multiple analog signals, heavily corrupted with additive white Gaussian noise. We improvise upon the MOSAICS architecture proposed by us in our previous work to include a wider class of signals having non-integral frequency components. This makes it possible to perform multiplexed compressed sensing for general frequency sparse signals.

Mapping joint ROM on a cube using electromagnetic trackers

This paper presents a unified framework using the unit cube for measurement, representation and usage of the range of motion (ROM) of body joints with multiple degrees of freedom (d.o.f) to be used for digital human models (DHM). Traditional goniometry needs skill and kn owledge; it is intrusive and has limited applicability for multi-d.o.f. joints. Measurements using motion capture systems often involve complicated mathematics which itself need validation. In this paper we use change of orientation as the measure of rotation; this definition does not require the identification of any fixed axis of rotation. A two-d.o.f. joint ROM can be represented as a Gaussian map. Spherical polygon representation of ROM, though popular, remains inaccurate, vulnerable due to singularities on parametric sphere and difficult to use for point classification. The unit cube representation overcomes these difficulties. In the work presented here, electromagnetic trackers have been effectively used for measuring the relative orientation of a body segment of interest with respect to another body segment. The orientation is then mapped on a surface gridded cube. As the body segment is moved, the grid cells visited are identified and visualized. Using the visual display as a feedback, the subject is instructed to cover as many grid cells as he can. In this way we get a connected patch of contiguous grid cells. The boundary of this patch represents the active ROM of the concerned joint. The tracker data is converted into the motion of a direction aligned with the axis of the segment and a rotation about this axis later on. The direction identifies the grid cells on the cube and rotation about the axis is represented as a range and visualized using color codes. Thus the present methodology provides a simple, intuitive and accura te determination and representation of up to 3 d.o.f. joints. Basic results are presented for the shoulder. The measurement scheme to be used for wrist and neck, and approach for estimation of the statistical distribution of ROM for a given population are also discussed.

A Novel Power Allocation Scheme for Two-User GMAC with Finite Input Constellations

Constellation Constrained (CC) capacity regions of two-user Gaussian Multiple Access Channels (GMAC) have been recently reported, wherein an appropriate angle of rotation between the constellations of the two users is shown to enlarge the CC capacity region. We refer to such a scheme as the Constellation Rotation (CR) scheme. In this paper, we propose a novel scheme called the Constellation Power Allocation (CPA) scheme, wherein the instantaneous transmit power of the two users are varied by maintaining their average power constraints. We show that the CPA scheme offers CC sum capacities equal (at low SNR values) or close (at high SNR values) to those offered by the CR scheme with reduced decoding complexity for QAM constellations. We study the robustness of the CPA scheme for random phase offsets in the channel and unequal average power constraints for the two users. With random phase offsets in the channel, we show that the CC sum capacity offered by the CPA scheme is more than the CR scheme at high SNR values. With unequal average power constraints, we show that the CPA scheme provides maximum gain when the power levels are close, and the advantage diminishes with the increase in the power difference.

Unique Type II Halogen center dot center dot center dot Halogen Interactions in Pentafluorophenyl-Appended 2,2 `-Bithiazoles

Herein, we report the design and synthesis of 2,2'-bithiazole derivatives with efficient intermolecular halogen interactions. The single crystal X-ray diffraction studies revealed unique type-II halogen interactions in these derivatives. The shortest type-II F center dot center dot center dot F interactions within the distance of 2.67 angstrom, at an angle of 89.1 degrees and 174.2 degrees, was observed for the first time. The Gaussian calculations were performed to further establish predominant F center dot center dot center dot F interactions.

An Outward-Wave-Favouring Finite Element-Based Strategy for Exterior Acoustical Problems

This work presents a finite element-based strategy for exterior acoustical problems based on an assumed pressure form that favours outgoing waves. The resulting governing equation, weak formulation, and finite element formulation are developed both for coupled and uncoupled problems. The developed elements are very similar to conventional elements in that they are based on the standard Galerkin variational formulation and use standard Lagrange interpolation functions and standard Gaussian quadrature. In addition and in contrast to wave envelope formulations and their extensions, the developed elements can be used in the immediate vicinity of the radiator/scatterer. The method is similar to the perfectly matched layer (PML) method in the sense that each layer of elements added around the radiator absorbs acoustical waves so that no boundary condition needs to be applied at the outermost boundary where the domain is truncated. By comparing against strategies such as the PML and wave-envelope methods, we show that the relative accuracy, both in the near and far-field results, is considerably higher.

Bilateral smoothing of gradient vector field and application to image segmentation

Medical image segmentation finds application in computer-aided diagnosis, computer-guided surgery, measuring tissue volumes, locating tumors, and pathologies. One approach to segmentation is to use active contours or snakes. Active contours start from an initialization (often manually specified) and are guided by image-dependent forces to the object boundary. Snakes may also be guided by gradient vector fields associated with an image. The first main result in this direction is that of Xu and Prince, who proposed the notion of gradient vector flow (GVF), which is computed iteratively. We propose a new formalism to compute the vector flow based on the notion of bilateral filtering of the gradient field associated with the edge map - we refer to it as the bilateral vector flow (BVF). The range kernel definition that we employ is different from the one employed in the standard Gaussian bilateral filter. The advantage of the BVF formalism is that smooth gradient vector flow fields with enhanced edge information can be computed noniteratively. The quality of image segmentation turned out to be on par with that obtained using the GVF and in some cases better than the GVF.