Removing noise from hyperspectral images can be very beneficial for improving classification accuracy. Recently, tensor robust principal component analysis (TRPCA) has been successfully employed to reduce noise in hyperspectral images. In TRPCA, a minimization involving a tensor nuclear norm and a$\ell _{1}$-norm is employed to separate the low-rank hyperspectral image from the sparse noise. Tensor nuclear norm minimization is solved by iteratively performing tensor singular-value thresholding (T-SVT). However, TRPCA possesses high computational complexity primarily due to the implementation of the T-SVT operator. The conventional approach for T-SVT is first to perform full tensor singular value decomposition (T-SVD) and then to shrink the singular values of the frontal slices in the frequency domain. However, when the solution of tensor nuclear norm minimization is a low-rank tensor, a good strategy is to incrementally find the singular values until they fall below the threshold. In this letter, we propose a randomized blocked algorithm for computing T-SVT, and we leverage the compression attained by the fast Fourier transform (FFT) to accelerate TRPCA. Numerical experiments indicate that our method is significantly faster than performing TRPCA via the full T-SVD while maintaining classification accuracy.

In recent years, prototypical networks have been widely used in many few-shot learning scenarios. However, as a metric-based learning method, their performance often degrades in the presence of bad or noisy embedded features, and outliers in support instances. In this paper, we introduce a hybrid attention module and combine it with prototypical networks for few-shot sound classification. This hybrid attention module consists of two blocks: a feature-level attention block, and an instance-level attention block. These two attention mechanism can highlight key embedded features and emphasize crucial support instances respectively. The performance of our model was evaluated using the ESC-50 dataset and the noiseESC-50 dataset. The model was trained in a 10-way 5-shot scenario and tested in four few-shot cases, namely 5-way 1-shot, 5-way 5-shot, 10-way 1-shot, and 10-way 5-shot. The results demonstrate that by adding the hybrid attention module, our model outperforms the baseline prototypical networks in all four scenarios.

Recently, many attention-based deep neural networks have emerged and achieved state-of-the-art performance in environmental sound classification. The essence of attention mechanism is assigning contribution weights on different parts of features, namely channels, spectral or spatial contents, and temporal frames. In this paper, we propose an effective convolutional neural network structure with a multichannel temporal attention (MCTA) block, which applies a temporal attention mechanism within each channel of the embedded features to extract channel-wise relevant temporal information. This multi-channel temporal attention structure will result in a distinct attention vector for each channel, which enables the network to fully exploit the relevant temporal information in different channels. The datasets used to test our model include ESC-50 and its subset ESC-10, along with development sets of DCASE 2018 and 2019. In our experiments, MCTA performed better than the single-channel temporal attention model and the non-attention model with the same number of parameters. Furthermore, we compared our model with some successful attention-based models and obtained competitive results with a relatively lighter network.

When a neural network is trained and used for inference in a digital hardware architecture, it is normally expected that the performance will translate exactly to new hardware. When implemented in low-power analog systems, there is no such luxury. This paper explores methods for reducing the impact of inter-device variations on the final performance of a neural network. An analysis of the effect of the network shape, including the number of layers and neurons in each layer, reveals a general pattern for minimizing the variability in accuracy. A number of means to reduce overfitting in neural networks are discussed and compared, and a more thorough demonstration is given to training methods that mitigate device overfitting.

Traditionally, the sliding window based activity recognition chain (ARC) has been dominating practical applications, in which features are carefully optimized towards scenario specifics. Recently, end-to-end, deep learning methods, that do not discriminate between representation learning and classifier optimization, have become very popular also for HAR using wearables, promising "out-of-the-box" modeling with superior recognition capabilities. In this paper, we revisit and analyze specifically the role feature representations play in HAR using wearables. In a systematic exploration we evaluate eight different feature extraction methods, including conventional heuristics and recent representation learning methods, and assess their capabilities for effective activity recognition on five benchmarks. Optimized feature learning integrated into the conventional ARC leads to comparable if not better recognition results as if using end-to-end learning methods, while at the same time offering practitioners more flexibility to optimize their systems towards specifics of wearables and their constraints and limitations.

Several architectures have been proposed for deep neural network (DNN)-based speech enhancement; however, these all utilize training targets related to the clean speech signal. In this paper, we evaluate the performance of several training targets in a noise-prediction DNN framework and compare the noise-prediction framework to a conventional speech-prediction network. Objective test results show that the mask-based targets are superior to the spectral magnitude target in the noise-prediction framework. The results also show that the best noise target outperforms the speech-prediction network in terms of objective quality and intelligibility metrics in seen noise conditions. The noise target is also competitive in unseen noise conditions, performing slightly worse in objective quality, but outperforming the speech-based target in objective intelligibility.

Researchers have recently examined a modified approach to sparse coding that encourages dictionaries to learn anomalous features. This is done by incorporating the matrix I-norm, or l1,∞ mixed matrix norm, into the dictionary update portion of a sparse coding algorithm. However, solving a matrix norm minimization problem in each iteration of the algorithm causes it to run more slowly. The purpose of this paper is to introduce block coordinate descent, a subgradient-like approach to minimizing the matrix norm, to the dictionary update. This approach removes the need to solve a convex optimization program in each iteration and dramatically reduces the time required to learn a dictionary. Importantly, the dictionary learned in this manner can still model anomalous features present in a dataset.

The extreme learning machine (ELM), an algorithm for training feedforward neural networks, is described in the literature as an algorithm that is suitable for both multiclass classification and multivariate regression problems. In this paper, we show that the closed-form ELM solution is not optimal for multivariate regression problems because it ignores correlations between the different response or target variables. We propose an improved algorithm, the canonical ELM, that accounts for the correlations between the target variables, and yet adheres to the ELM principle of learning without iteratively updating the weights in the network. Experimental results obtained using a synthetic dataset and several real-world datasets show that the canonical ELM has a higher prediction accuracy than the ELM and is also more stable.

Deep neural networks (DNNs) have recently been successfully applied to the speech enhancement task; however, the low signal-to-noise ratio (SNR) performance of DNN-based speech enhancement systems remains less than desirable. In this paper, we study an approach to DNN-based speech enhancement based on noise prediction. Three speech enhancement models based on noise prediction are proposed, and their performance is compared to that of conventional spectral-mapping models in seen and unseen noise tests. Objective test results show that the proposed noise prediction models perform well in enhancing speech quality in seen noise conditions and in enhancing high SNR speech signals. They also perform well in enhancing speech intelligibility in both seen and unseen noise conditions, but do not outperform the conventional models on quality metrics in unseen noise conditions. Further analysis of the enhanced speech signals is undertaken to explain the observed results.

In this paper, we propose a method for post-processing of deep neural network (DNN) enhanced speech. The method, which is simple and does not require additional training or expansion of the feature or target vectors, can be viewed as a mask-based approach in which a noisy speech signal is processed by a time-frequency (T-F) weighting derived from the noise-free spectral estimate of a DNN. A series of experiments and statistical analyses of results are carried out to compare the performance of the proposed approach to a baseline DNN enhancement system that features no post processing. Objective tests show that the proposed approach always improves both speech quality and intelligibility, and it outperforms a corresponding baseline system in both matched and mismatched noise conditions. Analysis of the enhanced speech shows that post processing reduces severe amplification distortions in the magnitude spectrum of the enhanced speech at the cost of a slight increase in severe attenuation distortions.

A frozen dictionary learning method is proposed to automatically separate the aspects of environmental audio that are different between normal data and anomalous data. The approach involves learning a dictionary-based sparse representation of the normal data, then freezing this portion of the dictionary while an added portion of the dictionary is learned on the anomalous data. Two dictionary learning algorithms are modified to allow training some elements while holding others constant. Both algorithms demonstrate the ability to separate the anomalies for two sets of test data. One set is chicken recordings with human crowd noise anomalies artificially added at -6 dB. The other set consists of recordings of chickens that are healthy and sick. Both dictionary methods are able to identify data anomalies with high accuracy and precision.

Physically and computationally efficient hardware coupled with fast sparse approximation solvers provide opportunities for real-time visual processing on low-power embedded platforms. This paper presents a system using the low-power Locally Competitive Algorithm (LCA) on the highly programmable, brain-inspired IBM TrueNorth chip. A small-scale spiking LCA network is successfully implemented on the TrueNorth chip, resulting in node dynamics comparable to that of a discretized LCA network. The sparse representations computed by the LCA implemented on TrueNorth result in minimal reconstruction error for every trial. This performance is achieved using only 11 of the available 4096 cores on the chip, offering the potential for scalability for real-world applications of this system.

It is well known that the variability in speech caused by the accents or dialects of speakers degrades the performance of speech recognition systems. One method to prevent this degradation is to correctly identify the accent or dialect of a speaker so that the putative system can be designed to use this information. In this paper, we apply the extreme learning machine, an efficient neural network learning algorithm, to the problem of accent/dialect classification on the TIMIT dataset. Mel frequency cepstrum coefficients (MFCCs) and the normalized energy parameter are computed from word samples in each dialect class, and these along with their first and second derivatives are used as training features. A novel architecture and a weighting scheme are used to classify North American dialects into seven groups. Using this technique, we obtain an classification accuracy of 76.92%, which to our knowledge is the best result reported for dialect classification on the TIMIT dataset.

In this work, for a given a set of code vector assignments to an input by a multistage residual vector quantizer RVQ [1], Bayesian framework is formulated to find the most probable class membership of the input. Furthermore, Markov structure is also used to improve the memory cost of the classification.

This paper proposes a novel speaker adaptation algorithm for classifying speech based on deep neural networks (DNNs). The adaptation algorithm consists of two steps. In the first step a deep neural network is trained using raw Mel-frequency cepstral coefficient (MFCC) features to discover hidden structures in the data and employing the activations of the last hidden layers of the DNN as acoustic features. In the second step using nearest neighbor, an adaptation algorithm learns speaker similarity scores based on a small amount of adaptation data from each target speaker using the DNN-based acoustic features. Based on the speaker similarity score, classification is done using a k-nearest neighbor (k-NN) classifier. The novelty of this work is that instead of modifying and re-training the DNN for speaker adaptation, which comprises a large number of parameters and is computationally expensive, activations of the learned DNN are used to project features from MFCC to a sparse DNN space, then speaker adaptation is performed based on similarity (i.e. nearest neighbor) using k-NN algorithm. With only a small amount of adaptation data, it reduces the number of phoneme classification error in the TIMIT dataset by 23%. This work also analyzes impact of deep neural networks architecture on speaker adaptation performance.

Speech recognition systems are either based on parametric approach or non-parametric approach. Parametric based systems such as HMMs have been the dominant technology for speech recognition in the past decade. Despite a lot of advancements and enhancements in the design of these systems: key problems such as long term temporal dependence, etc. Has not yet been solved. Recently due to availability of large amount of data and cheap computing resources (processing power and memory) non-parametric based approach to solve speech recognition and classification task is becoming popular and feasible. The key advantage of non-parametric based approach is that all the information from the training data is retained as we don't approximate our data with specific statistical models resulting in more speaker specific information. In this paper we propose a k-nearest neighbor (k-NN) phoneme prediction scheme using speaker ranking vector. Speaker ranking vector is calculated by finding the similarity of the given TEST speaker with the Instance Space using k-NN. The results were compared with nearest neighbor and k-NN majority voting approach. Our proposed scheme gives a better prediction accuracy as compare with nearest neighbor and k-NN majority voting scheme. This approach can help speech recognizer to customize on the fly for a given talker and customize training data on the basis of similarity measure. In this preliminary research we are using a small amount of data to train our phoneme prediction classifier engine. Performance can be further increase by increasing the training data for finding speaker ranking.

We describe architectures for audio classification front ends on a reconfigurable analog platform. Real-time implementation of audio processing algorithms involving discrete-time signals tend to be power-intensive. We present an alternate continuous-time system implementation of a noise-suppression algorithm on our reconfigurable chip, while detailing the design considerations. We also describe a framework that enables future implementations of other speech processing algorithms, classifier front ends, and hearing aids.

In this paper, a reordering vector quantization algorithm based on conditional probabilities of a codebook transition matrix is implemented. The dynamic reordering of the codebook dramatically decreases the entropy for temporally structured sources, enabling a second stage of entropy coding to further improve the efficiency of VQ. Results from synthetic, Markov sources and speech sources are shown to outperform other baseline vector quantization algorithms.

While most single-channel noise reduction algorithms fail to improve speech intelligibility, the ideal binary mask (IBM) has demonstrated substantial intelligibility improvements. However, this approach exploits oracle knowledge. The main objective of this paper is to introduce a novel binary mask estimator based on a simple sparse approximation algorithm. Our approach does not require oracle knowledge and instead uses knowledge of speech structure.

Dual-tree algorithms are a widely used class of branch-and-bound algorithms. Unfortunately, developing dual-tree algorithms for use with different trees and problems is often complex and burdensome. We introduce a four-part logical split: the tree, the traversal, the point-to-point base case, and the pruning rule. We provide a meta-algorithm which allows development of dual-tree algorithms in a tree-independent manner and easy extension to entirely new types of trees. Representations are provided for five common algorithms; for k-nearest neighbor search, this leads to a novel, tighter pruning bound. The meta-algorithm also allows straightforward extensions to massively parallel settings.

Online music lessons are growing more and more prevalent with resources such as Massive Open Online Courses (MOOCs) and lessons via YouTube videos. We discuss two issues in online music lessons that machine learning techniques may be used to solve: finding or developing a music lesson based on the student's learning style, musical background, or preference and quantitative assessment of the student's performance (whether as an online-instructor or self-taught through online videos). In particular, we discuss two solutions with specific applications. First, we propose a method for automating music lessons for learning how to play pre-recorded songs via existing music information retrieval (MIR) techniques for adaptability to various learning styles. Secondly, we propose a method for automating the assessment of rhythmic structures (such as tremolo, vibrato, etc.) via quantitative metrics of comparing modulation spectral features. We then list some resources and software that are currently available to integrate these methods to enhance online music education.

Many developers of audio signal processing strategies rely on objective measures of quality for initial evaluations of algorithms. As such, objective measures should be robust, and they should be able to predict quality accurately regardless of the dataset or testing conditions. Kates and Arehart have developed the Hearing Aid Speech Quality Index (HASQI) to predict the effects of noise, nonlinear distortion, and linear filtering on speech quality for both normal-hearing and hearing-impaired listeners, and they report very high performance with their training and testing datasets [Kates, J. and Arehart, K., Audio Eng. Soc., 58(5), 363-381 (2010)]. In order to investigate the generalizability of HASQI, we test its ability to predict normal-hearing listeners' subjective quality ratings of a dataset on which it was not trained. This dataset is designed specifically to contain a wide range of distortions introduced by real-world noises which have been processed by some of the most common noise suppression algorithms in hearing aids. We show that HASQI achieves prediction performance comparable to the Perceptual Evaluation of Speech Quality (PESQ), the standard for objective measures of quality, as well as some of the other measures in the literature. Furthermore, we identify areas of weakness and show that training can improve quantitative prediction.

Statistical analysis of computations per unit energy in processors over the last 30 years is given that illustrates a sharp reduction in the rate of energy efficiency improvements over the last several years resulting in the formation of an asymptotic “wall” with our dataset; we use the measure of giga multiply accumulates per Joule. We have developed an energy model which takes into account the realities of scaling, specifically for asynchronous systems. Studies of an energy efficient asynchronous pipeline show fabricated results of 17 Giga Operations per Joule in 0.6 μm at subthreshold when fully pipelined, and simulations at a more modern 65 nm process show a further order of magnitude improvement on that.

Different noise-reduction methods have been proposed in the literature for single and multiple-microphone applications. For binaural hearing aids, multiple-microphone noise-reduction methods offer two significant psycho-acoustical advantages with respect to the single-microphone methods. First, noise reduction strategies based on binaural processing (processing using the information received at the left and right ear) are more effective than using independent monaural processing because of the added information. Second, there is an user preference for noise reduction methods that preserve localization cues of both target and interfering signals. Although different multiple-microphone noise-reduction techniques have been proposed in the literature, only a small set is able to preserve the localization cues for both target and interfering signals. This paper proposes a binaural noise-reduction method that preserves the localization cues for both target and interfering signals. The proposed method is based on blind source separation (BSS) followed by a postprocessing technique inspired by a human auditory model. The performance of the proposed method is analyzed using objective and subjective measurements, and compared to existing binaural noise-reduction methods based on BSS and multichannel Wiener filter (MWF). Results show that for some scenarios and conditions, the proposed method outperforms on average the existing methods in terms of noise reduction and provides nearly similar sound quality.

For multiple-input multiple-output communication systems that employ four or more transmit and four or more receive antennas, symbol detection remains a challenge in communication systems research. Lattice-reduction-aided detectors are attractive solutions to this problem because these detectors achieve the same diversity as the maximum-likelihood detector while exhibiting lower complexity. Current applications of lattice-reduction-aided detectors involve executing a lattice reduction algorithm to completion and then utilizing this result in the subsequent symbol detection. In this article, however, we examine the possibility of partially executing the lattice reduction algorithm. We first demonstrate using a hypothetical lattice-reduction-aided detector that early termination of lattice reduction algorithms is possible in the context of MIMO detection. Encouraged by these results, we develop and introduce incremental lattice reduction, which utilizes a practical early termination condition. We then apply this idea to develop a joint symbol detection and lattice reduction algorithm that is based on the Lenstra, Lenstra, Lovasz algorithm and successive interference cancellation. An evaluation using a spatial correlation channel model demonstrates that the proposed algorithm effectively distributes the lattice reduction processing over the length of each received packet. This behavior naturally enables the relaxation of throughput and latency requirements of lattice reduction algorithm hardware realizations.

Power management in loudspeakers is used to extend the battery life of portable electronics and to prevent thermal damage to the speakers. Traditionally, speaker power management has been implemented as an adaptive gain control, which limits the power applied to the loudspeaker uniformly. In this pa per, we present a perception-based algorithm that suppresses inaudible contents from the audio before sending it to the amplifier and loudspeaker. The power reduction algorithm attenuates the signal energy in the digital domain. Subjective test results show that the energy reduction is largely unnoticeable in small loudspeakers.

The question in the title is relevant when considering lattice-reduction-aided MIMO detectors, which achieve the same diversity as the maximum-likelihood detector while exhibiting lower complexity. In this paper we examine if all basis updates, which account for the largest complexity contribution in the Lenstra, Lenstra, Lovasz lattice reduction algorithm, are necessary for lattice reduction in the context of MIMO detection. We first provide an abstract answer to this question in the form of an idealized experiment that demonstrates the potential for a large reduction in the number of basis updates even when spatial correlation is present. Encouraged by these results, we seek a practical answer to this question by formulating a joint lattice reduction and symbol detection algorithm based on successive interference cancellation. Experimental results of the proposed method demonstrate that on average only 10% to 25% of basis updates are necessary on average depending on the degree of spatial correlation. Therefore, the answer to the question in the title is an encouraging no.

Commonly used speech enhancement algorithms estimate the power spectral density of the noise to be removed, or make a decision about the presence of speech in a particular frame, and estimate the clean speech based on these. Errors in a noise estimate or speech activity decision may result in undesirable artifacts, and some errors may be more damaging than others. Robust Bayesian Analysis is used to analyze the sensitivity of algorithms to errors in noise estimates and improve signal-to-noise ratio while mitigating artifacts in the enhanced speech. The findings explain why some common heuristic changes to the Wiener filter algorithm are effective. A standard Wiener algorithm is used for comparison, objective quality measures are used to quantify improvement, and insights into the underlying mechanisms of heuristic methods are offered.

The design of analog classifiers constitutes a trade- off between performance and complexity, and designers have historically adopted more complex architectures to lower the error rate of a classification task. An alternative design paradigm is presented in this paper: We design the front-end of a sound classification system with simple "base" classifiers. We then enhance the overall performance with the aid of the AdaBoost algorithm, which selects the most appropriate "base" classifiers and combines them with different weights. We describe the general architecture and the algorithm to select features and present a design example with simulation results in a TSMC- compatible 0.35-μm technology.

Most speech enhancement applications perform frequency shaping by means of multiplication in the frequency domain. Operating in the frequency domain is equivalent to convolution in the time domain. In these speech enhancement algorithms, the updating of frequency response alone cannot ensure the fulfillment of the conditions required for multiplication in frequency to correspond to linear convolution instead of circular convolution. As a result, artifacts and distortions may be present in the output of a standard fast Fourier transform (FFT)-based algorithm. Typical methods to deal with these artifacts involve overlapping and windowing. However, even using these strategies, artifacts may be perceptually noticeable under certain signal-to-noise ratio (SNR) conditions and/or when a high sampling frequency is employed. This paper analyzes the efficiency of the standard methods, explains the source of these distortions, provides a perceptual evidence of these artifacts, and proposes two alternative methods to perform artifact-free and distortion-free FFT convolution. These methods are based on the extension of the impulse response and the splitting of the impulse response in two impulse responses, operations that are performed in the frequency-domain. Computational costs and performance of the proposed techniques are also discussed.

When manipulating audio signals, frequency decomposition, gain functions, and temporal behavior can be understood in terms of modulation effects and their perceptual impact. For example, audio enhancement or dynamic range control both rely on spectral gain functions that fluctuate rapidly enough to adapt to changing signal characteristics. However, rapid fluctuations in gain effectively modulate the signal, resulting in perceptual artifacts. This paper discusses the perceptual impact of modulating gain and trade-off between temporal responsiveness and perceptibility. We demonstrate that, in many situations, it is best to apply gains in subbands that are similar to critical bands in the human auditory system and that the gain function in each subband should be low-pass filtered with a cutoff frequency proportional to the subband bandwidth.

In most speech enhancement algorithms, a frequency response is computed per each block, and a multiplication in the frequency domain is used to apply this frequency response. Typically, in these algorithms, the updating procedure cannot ensure the fulfillment of the condition to perform a linear convolution instead of a circular convolution. As a result, artifacts and distortions may be present. Standard methods to deal with these distortions use overlapping and windowing. This work presents a detailed analysis of these distortions and proposes a new method to perform a artifact-free and distortion-free convolution in the FFT domain.

For cell-phone applications, single microphone noise suppression techniques have limited performance at very low SNR (close to 0 dB). In certain cases, they also suffer from the artifacts of nonlinear processing. In this paper, we will show that techniques based on two-microphone blind source separation (BSS) algorithm provide significant interference suppression for cell-phone applications, particularly at low operating SNR values. We also propose a post-BSS processing method based on frequency-domain spectral subtraction that further improves the BSS speech output in diffused noise case. We optimize the BSS algorithm so that it can be implemented on a low-power audio codec processor, which would be ideal for cell phone applications. Furthermore, based on extensive analysis under different noise and acoustic conditions, we suggest recommendations for optimal placement of microphones on a cell phone. We also study the trade-off between the unmixing filter length and the noise suppression performance. In all our experiments, we use real recordings made on a cell phone equipped with two microphones.

Lattice reduction-aided equalization techniques have emerged as a low-complexity method to achieve the same diversity as maximum likelihood detectors. We address the VLSI implementation of these LR-aided equalizers by modifying the CLLL algorithm from a fixed-point hardware perspective. We then apply the modified algorithm together with additional micro-architecture and operation scheduling enhancements to create an updated CLLL processor. Finally, through BER simulations and FPGA synthesis results we demonstrate the suitability of our CLLL processor for integration into a 64-QAM MIMO system.

An efficient way for computing the response of an adaptive digital filter is to use sliding-block distributed arithmetic (SBDA). One disadvantage of distributed arithmetic is the amount of memory utilized. By encoding the memory tables in offset binary code (OBC), the size of the memory tables is reduced in half. However, the computational workload remains unchanged. By modifying the computational flow, the computational workload can be reduced by almost half at the expense of slightly more memory. This modified SBDA structure is called SBDA-OBC. It has memory requirements 25%-50% lower than SBDA depending on the size of the sub-filter. In terms of the computational workload, SBDA-OBC is most advantageous for large sub-filters and when the filter is split into few subfilters. In this case, the computational workload is reduced almost in half.

This paper presents a high-throughput, low-latency, hardware-efficient fixed-point implementation of the fast affine projection (FAP) algorithm. The proposed architecture utilizes reusable distributed arithmetic (RDA) in combination with optimizations in the update process to enable the coefficients to be updated in a fixed number of cycles independent of filter length. Fixed-point simulations show that the effect of replacing some of the multiplications with arithmetic shifts is minor, with RDA-FAP maintaining a faster convergence rate than NLMS. The proposed design is also compared against a multiplier-based design in terms of number of computations and number of clock cycles needed for a single FAP update cycle.

In the current work, a novel analog hardware to denoise speech signals has been proposed and simulated in 0.5µm VLSI technology. The system is based on Automatic Gain Control which suppresses the noisy parts while boosting clean intervals of a signal. Further, the architecture is devised using Operational Transconductance Amplifiers (OTA) operating in sub-threshold region for attaining high programmability, low power and also to facilitate testing on an RASP 2.7 FPAA Prototype. A test noisy signal of 5dB SNR when applied to the system resulted in a 13.5dB improved output SNR consuming 0.53mW power from a 1.6V supply.

Motivated by the unwillingness to accept the worst-case timing constraint that synchronous logic imposes, and additionally motivated by finding a supply voltage scaling scheme for datapath circuits that is unconstrained by timing errors in memory elements, the authors have built an asynchronous datapath that is embedded seamlessly into a synchronous register file. This paper will show that not only does asynchronous arithmetic logic exhibit many characteristics that allow it to be inherently lower power, but it is significantly faster than any synchronous counterpart and is a perfect candidate technology for datapath acceleration. Further, novel circuits are presented that allow asynchronous datapath units to be embedded in a synchronous environment with little overhead while the dual-rail asynchronous encoding scheme is successfully converted with equally low overhead. The authors have built a test chip being fabricated at the time of publication. The circuits on this chip will be discussed and simulation results given showing this design to be both energy and performance efficient when compared to other known datapath designs.

In this paper we explore avenues for improving the reliability of dimensionality reduction methods such as Non-Negative Matrix Factorization (NMF) as interpretive exploratory data analysis tools. We first explore the difficulties of the optimization problem underlying NMF, showing for the first time that non-trivial NMF solutions always exist and that the optimization problem is actually convex, by using the theory of Completely Positive Factorization. We subsequently explore four novel approaches to finding globally-optimal NMF solutions using various ideas from convex optimization. We then develop a new method, isometric NMF (isoNMF), which preserves non-negativity while also providing an isometric embedding, simultaneously achieving two properties which are helpful for interpretation. Though it results in a more difficult optimization problem, we show experimentally that the resulting method is scalable and even achieves more compact spectra than standard NMF.

In this paper, we propose a temporal dynamic index reordering vector quantization for wavelet-based embedded coding. Da Silva et al. introduced a vector quantization concept that is similar to EZW called a successive approximation vector quantizer (SAVQ). The successive refinement process is defined as a temporal process in our proposed algorithm. The temporal updates are performed in every refinement pass, and the updates of codevectors reflect the updates of angles for vector approximation. Because the approximation trajectory of SAVQ is similar to that of the least-mean- square algorithm, temporal redundancy does not seem to be obvious. However, the redundancy becomes more clear in the improved SAVQ. By carefully analyzing the angle transitions, we are successfully able to apply dynamic index reordering vector quantization (DIRVQ) in the temporal domain and improve the coding performance. Efficient encoding and decoding algorithms for D4lattice vector quantization are also proposed, and the algorithms can be applicable to DIRVQ to reduce the computational complexity of the reordering process. Experiments are performed with the 2x2 vector size. For almost all tested bpp's the PSNR performance of our proposed algorithm for a lena image outperforms that of SAVQ up to 0.18 dB. Improvement is measured for the highly detailed baboon image as well.

In this paper, we propose a spatially recursive optical flow estimation (OFE) framework using adaptive filtering. One of most successful OFE algorithms is a gradient-based least- squares (LS) within a local image window because of high performance and low-complexity. However, it has some redundancies for calculating successive LS among adjacent pixels. Therefore, we suggest an efficient framework using recursive least-squares (RLS) and adaptive filtering to improve the computational efficiency. The performance and computational complexity are compared to least-squares OFE and spatially recursive OFE algorithms. Based on these results, we conclude that our proposed algorithm framework under proper window size can reduce computational complexity especially as the number of motion modeling parameters increases by using the property of RLS and adaptive filtering.

A 16 × 16 programmable analog radial-basis-function (RBF) based classifier is demonstrated. The distribution of each feature is modeled by a Gaussian function, which is realized by a proposed floating-gate bump circuit having bell-shaped transfer characteristics. The maximum likelihood, mean, and variance of the distribution are stored in floating-gate transistors and are independently programmable. By cascading these floating-gate bump circuits, the overall transfer characteristics approximate a multivariate Gaussian distribution with a diagonal covariance matrix. An array of these circuits constitutes a compact RBF-based classifier. When followed by a winner-take-all circuit, the analog classifier can implement vector quantization. Automatic gender identification is implemented on a 16 × 16 analog vector quantizer chip as one possible audio application of this work. The performance of the analog classifier is comparable to that of digital counterparts. The proposed approach can be at least two orders of magnitude more power efficient than the digital microprocessors at the same task.

This paper discusses the application of a computational image sensor, capable of performing separable 2-D transforms on images in the analog domain, to compressive sensing. Instead of sensing and transmitting raw pixel data, this image sensor first projects the image onto a separable 2-D basis set. The inner products computed in these projections are computed in the analog domain using a computational focal-plane and a computational analog vector-matrix multiplier. Since this operation is performed in the analog domain, components such as the analog-to-digital converters can be taxed less when a only subset of correlations are performed. Compressed sensing theory prescribes the use of a pseudo-random, incomplete basis set, allowing for sampling at less than the Nyquist rate. This can reduce power consumption or increase frame rate.

Modern advances in reconfigurable analog technologies are allowing field-programmable analog arrays (FPAAs) to dramatically grow in size, flexibility, and usefulness. This paper presents rapid prototyping results of a bandpass filter as a sample analog circuit using our floating-gate based large-scale FPAA. A major source of parasitics introduced during the circuit mapping process is interconnect switches used for routing. Our goal is to obtain models of the mapped circuits that can be simulated using SPICE in order to observe the impact of interconnect parasitics on the relevant analog metrics. Our results indicate that the mapped analog circuits obtain desired responses even with interconnect parasitics, clearly demonstrating the practicality of our FPAA.

Previously we showed that band-pass filtered MFCC-like features are useful for noise robust speech discrimination and recognition. In this paper we aim to improve the previously presented features by incorporating varying time constants and gain adaptation in each frequency channel. We show that varying the time constants leads to a representation that is less prone to the effects of noise. Further, we show that gain adaptation can not only provide better performance in clean condition but can also be used to improve the noise robustness of the features. These improvements come at a very small increase in computational cost. Speech discrimination and recognition results are presented.

Comb filters are popular in the speech enhancement field because of their ability to suppress noise in the voiced speech spectrum without degrading the speech quality. Adaptive comb filters (ACF) can suppress significant amounts of noise in voiced speech using the quasiperiodic properties of the voiced speech sounds. However, they rely on accurate voicing and pitch epoch detection, which is challenging at low SNRs and in nonstationary noise conditions. Here, we propose using a non-acoustic auxiliary sensor for detecting the pitch epochs and voicing. Experiment results with the ACF filter showed practically no noise reduction when noisy speech is used for pitch epoch and voicing detection. However, when the auxiliary sensor signal is used, significant improvements are obtained. Therefore, it is shown that the ACF system becomes useful, at least under the considered noise conditions, only if the auxiliary data is available. In addition to the dramatic gain with the auxiliary sensors, performance of the ACF system is further boosted by using it in tandem with a frequency-domain ACF system proposed here. Objective measures, spectrogram analysis, and subjective listening test results clearly show substantial improvement with the tandem system compared to the time-domain ACF system.

We present a viewpoint showing that analog signal processing approaches are becoming configurable and programmable like their digital counterparts, while retaining a huge computational efficiency, for a given power budget, compared to their digital counterparts. We present recent results in programmable and configurable analog signal processing describing the widespread potential of these approaches. We also discuss issues with configurable systems, including size, power, and computational tradeoffs, as well as address the computational efficiency of these approaches.

Feature fusion is a popular approach for improving the accuracy of speech recognition systems in noisy environments. Although the feature fusion method performs well for high and moderate SNRs, its performance degrades rapidly at low SNRs. Moreover, auxiliary features that are robust to acoustic noise may sometimes be unreliable, because of sensor misplacements etc, which can further degrade the performance. Furthermore, noisy sensor signals may exist not only in the test data but also in the training data. Here, the feature fusion method is combined with a missing data technique to improve noise-robustness at low SNRs. An auxiliary feature is used both for feature fusion and for detecting the unreliable speech frames. Noisy auxiliary features are addressed using a missing data approach both in the decoding and training systems. In the experiments, substantial improvements in the feature fusion method are obtained especially at low SNRs. The proposed missing data-based training strategy is also shown to improve the accuracy significantly.

The paper proposed a new glass break detector analog front-end to replace simple signal threshold or signal slope threshold methods that normally trigger modern power-intensive DSP-based glass break detectors. The new front-end extracts features related to the total signal energy, high-band energy, and zero-crossings using analog circuits and inputs these into a novel analog classifier circuit that determines if a given 20 ms window of sound resulted from glass breaking. The paper presents TSMC 0.35μm SPICE system simulation results of the front-end, using actual sound files from a sound database to drive the simulations. Through these simulations the authors shows the utility of using analog circuits for both feature extraction and classification. Using a developmental database of glass break sounds and non-glass break sounds that includes the categories of noise, music, and speech, the analog front-end has 100% testing accuracy.

We present an implementation of a reconfigurable16-tapfinite impulse response filter for post-processing applications. This filter exploits the distributed arithmetic technique for signal processing and floating-gate voltage references for setting tunable analog coefficients. The filter is fabricated in0.5μmCMOS process, and its order can be increased at the cost of0.011mm2of die area and0.02mWof power per tap. Measurement results for low-pass and band-pass filters at50kHzsampling frequency are presented.

A 2-MITE is a multiple-input translinear element with two input gates. In this paper, different properties of networks of 2-MITEs are derived, especially in the case of product-of-power-law (POPL) networks, in which the output currents are products of the inputs raised to different powers. It is found that conditions ensuring the uniqueness and stability of the operating point in 2-MITE networks are less stringent than those for MITE networks with higher number of input gates. This simplifies the synthesis of these networks considerably. A graph-theoretic approach to the analysis of 2-MITE networks is presented. Necessary conditions for a set of power-law equations to be implementable by 2-MITE networks are derived. Sufficient conditions for the same are presented for the case of POPL networks with one output

A procedure for synthesizing multiple-input translinear element (MITE) networks that implement a given system of translinear-loop equations (STLE) is presented. The minimum number of MITEs required for implementing the STLE, which is equal to the number of current variables in the STLE, is attained. The number of input gates of the MITEs is minimal amongst those MITE networks that satisfy the STLE and have the minimum number of MITEs. The synthesized MITE networks have a unique operating point and, in many cases, the network is guaranteed to be stable in a particular sense. This synthesis procedure exploits the relationship between MITE product-of-power-law (POPL) networks and linear diophantine equations which is explored in detail here.

In this paper, we propose a new algorithm for wavelet-based multi-resolution motion estimation (MRME) using temporal aliasing detection (TAD). In wavelet transformed image/video signals, temporal aliasing will be severe as the motion of object increases, causing the performance of the conventional MRME algorithms to drop. To overcome this problem, we perform the temporal aliasing detection and MRME simultaneously instead of using a temporal anti-aliasing filter which changes the original signal. We show that this technique gives competitive or better performance in terms of rate-distortion (RD) for slow-varying or simple-moving video signals compared to conventional MRME employing increased search area (SA).

A compact analog programmable multi-dimensional radial basis function (RBF) based classifier is demonstrated. The probability distribution of each feature in the templates modeled by a Gaussian function is approximately realized by the transfer characteristics of a floating-gate bump circuit. The maximum likelihood, the mean, and the variance can be inde- pendently programmed. By cascading these floating-gate bump circuits, the transfer characteristics approximate a multivariate Gaussian function with a diagonal covariance matrix. An array of these circuits constitute a compact multi-dimensional RBF- based classifier. When followed by a winner-take-all circuit, the RBF-based classifier forms an analog vector quantizer. We use receiver operating characteristic curves and equal error rate to evaluate the performance of our analog classifiers. We show that the analog classifier performance is comparable to that of digital counterparts. The proposed approach can be at least two orders of magnitude more power efficient than the digital microprocessors at the same task.

In this paper, a speech signal estimation framework involving Kalman filters for use as a front-end to the Aurora-2 speech recognition task is presented. Kalman-filter based speech estimation algorithms assume autoregressive (AR) models for the speech and the noise signals. In this paper, the parameters of the AR models are estimated using a expectation-maximization approach. The key to the success of the proposed algorithm is the constraint on the AR model parameters corresponding to the speech signal to belong to a codebook trained on AR parameters obtained from clean speech signals. Aurora-2 noise-robust speech recognition experiments are performed to demonstrate the success of the codebook-constrained Kalman filter in improving speech recognition accuracy in noisy environments. Results with both clean and multi-conditional training are provided to show the improvements in the recognition accuracy compared to the base-line system where no pre-processing is employed

In this paper, we propose a new architecture for hierarchical block motion estimation (HBME) algorithms using variable block sizes. The binary tree architecture (BTA) is well suited to HBME with constant block size because of the interdependence between computations in different levels of search. However, HBME algorithms based on variable block sizes can decrease the processing element (PE) utilization of BTA. The modified binary tree architecture (MBTA) presented here improves the PE utilization and area efficiency of BTA with low overhead for HBME algorithms using variable block sizes. We describe the architecture of MBTA and present the comparison of the performance with BTA for wavelet-based multi-resolution motion estimation (MRME) algorithms.

In this paper, we propose a checkerboard-type filtering algorithm for a low-power gradient-based optical flow estimation (OFE) system. A gradient-based optical flow algorithm estimates optical flow field using gradient values of images. The conventional filtering stages in OFE that are used to calculate gradient values are composed of a 3D prefiltering/smoothing stage and three 1D parallel derivative filtering stages. To reduce power consumption using pixel-wise parallel processing in analog spatial transform imager, we suggest a checkerboard-type filtering based on convolution theorem and common data sharing. We describe the equivalence of our filtering scheme with the conventional smoothing and derivative filtering schemes and present some comparison in terms of the number of operations and power savings. We show that our approach is promising for low-power implementation of gradient-based OFE system under proper data scanning direction.

This paper presents the performance test and analysis of a correlation-based optical flow estimation (CBOFE) algorithm, the Camus method, under various noisy environments. The noise sensitivity of the original algorithm is tested under different noise sources such as additive Gaussian, bit quantization multiplicative noises and combinations of multiplicative and additive noises. We show the efficiency of pre-filtering under different noisy environments. The SNR equations are derived for various additive and multiplicative noise cases for analyzing the performance degradation. The simulation results show that the multiplicative noise is comparable to additive noise if the standard deviations of noises are scaled to the same percentage of the full range of signal.

Modern advances in reconfigurable analog technologies are allowing field-programmable analog arrays (FPAAs) to dramatically grow in size, flexibility, and usefulness. Our goal in this paper is to develop the first placement algorithm for large-scale floating-gate-based FPAAs with a focus on the minimization of the parasitic effects on interconnects under various device-related constraints. Our FPAA clustering algorithm first groups analog components into a set of clusters so that the total number of routing switches used is minimized and all IO paths are balanced in terms of routing switches used. Our FPAA placement algorithm then maps each cluster to a computational analog block (CAB) of the target FPAA while focusing on routing switch usage and balance again. Experimental results demonstrate the effectiveness of our approach

Modern advances in reconfigurable analog technologies are allowing field-programmable analog arrays (FPAAs) to dramatically grow in size, flexibility, and usefulness. Our goal in this paper is to develop the first hierarchical placement algorithm for large-scale floating-gate based FPAAs with a focus on the minimization of the parasitic effects on interconnects under various device-related constraints. In our hierarchical approach, our FPAA clustering algorithm first groups analog components into a set of clusters so that the total number of routing switches used is minimized and all IO paths are balanced in terms of routing switches used. Our FPAA placement algorithm then maps each cluster to a computational analog block (CAB) of the target FPAA while focusing on routing switch usage again. Experimental results demonstrate the effectiveness of our approach.

Intelligibility of spoken words in noisy environments is an important problem of speech coders particularly for military applications. The intelligibility problem of a MELP speech encoder in noisy environments is addressed by using a novel speech enhancement algorithm at the front end. The speech signal is segmented into broad phonetic classes using auxiliary sensors in addition to the acoustic microphone. Each phoneme class is enhanced by suppressing maximum noise while minimally distorting perceptually important cues using the acoustic-phonetic knowledge about the class. The DRT (diagnostic rhyme tests) cores in an M2 tank noise environment show substantial improvement over the MELPe coder.

The paper presents a new memory-efficient distributed arithmetic (DA) architecture for high-order FIR filters. The proposed architecture is based on a memory reduction technique for DA look-up-tables (LUTs); it requires fewer transistors for high-order filters than original LUT-based DA, DA-offset binary coding (DA-OBC), and the LUT-less DA-OBC. Recursive iteration of the memory reduction technique significantly increases the maximum number of filter order implementable on an FPGA platform by not only saving transistor counts, but also balancing hardware usage between logic element (LE) and memory. FPGA implementation results confirm that the proposed DA architecture can implement a 1024-tap FIR filter with significantly smaller area usage (<50%) than the original LUT-based DA and the LUT-less DA-OBC.

Modern advances in reconfigurable analog technologies are allowing field-programmable analog arrays (FPAAs) to dramatically grow in size, flexibility, and usefulness. With these advances, analog circuits and systems can be programmable, reconfigurable, adaptive, implemented on standard CMOS to take advantage of scaled CMOS technology, and at a density comparable to digital memories. Our goal in this paper is to develop the first physical design automation toolset for floating-gate based FPAA with focus on minimization of parasitic effects on FPAA interconnect. We provide graph-based analog circuit and FPAA device modeling suitable for efficient mapping. Our FPAA clustering algorithm constructs Computational Analog Blocks (CAB) from analog circuit elements while improving the utilization of the device and reducing its impact on the total number of routing switches used. Experimental results demonstrate the effectiveness of our approach.

In this paper, an FIR adaptive filter implementation using a multiplier-free architecture is presented. The implementation is based on distributed arithmetic (DA) which substitutes multiply-and-accumulate operations with a series of look-up-table (LUT) accesses. This can be achieved at the cost of a moderate increase in memory usage. The proposed design performs an LMS-type adaptation on a sample-by-sample basis. This is accomplished by an innovative LUT update using a matched auxiliary LUT. The system is implemented on an FPGA that enables rapid prototyping of digital circuits. Implementation results are provided to demonstrate that a high-speed LMS adaptive filter can be realized employing the proposed architecture.

In this paper, an FIR adaptive filter implementation, using a multiplier-free architecture, is presented. The implementation is based on distributed arithmetic (DA) which substitutes multiply-and-accumulate operations with a series of look-up-table (LUT) accesses. This can be achieved at the cost of a moderate increase in memory usage. The proposed design performs an LMS-type adaptation on a sample-by-sample basis. This is accomplished by an innovative LUT update using a matched auxiliary LUT. The system is implemented on an FPGA that enables rapid prototyping of digital circuits. Implementation results are provided to demonstrate that a high-speed, low logic complexity LMS adaptive filter can be realized employing the proposed architecture.

In the past researchers have proposed a variety of features that are based on the human auditory system. However none of these features have been able to replace mel-frequency cepstral coefficients (MFCC) as the preferred feature for audio classification problems, either because of computational costs involved or because of their poor performance in the presence of noise. In this paper we present new features derived from a model of the early auditory system. We compare the performance of the new features with MFCC in a four-class audio classification problem and show that they perform better. We also test the noise robustness of the new features in a two-way audio classification problem and show that it outperforms the MFCC. Further, these new features can be implemented in low-power analog VLSI circuitry making them ideal for low-power sensor networks.

A voicing feature is used in concatenation to MFCC features to increase the performance of digit recognition at both low and high SNRs. The problem of noise robust extraction of the voicing feature is solved by using the glottal electromagnetic sensor (GEMS). The GEMS device provides reliable voicing information at all SNRs and noise environments. It is shown that although the voicing feature increases the performance for the clean speech case, the relative improvement for the noisy case is significantly higher for a digit recognition task. Our results indicate that the GEMS device can solve the fundamental problem of extracting reliable voicing information in noisy environments.

A soft-decision Ephraim-Malah suppression rule based speech enhancement algorithm is proposed for intelligibility enhancement in parametric speech coders. A glottal sensor is used to improve the intelligibility of a baseline system that uses only the acoustic microphone. The objective measure test shows that the proposed system decreases the spectral distortion by 2-3 dB for most phonetic classes. Moreover, significant improvements in DRT scores of nasality and sibilation features are obtained compared to the baseline system when the noise suppression systems are concatenated with a MELP based speech coder.

This paper describes a single acoustic–channel speech enhancement, utilizing an auxiliary non-acoustic sensor. Unlike classical algorithms, which make use of the knowledge from acoustic signal alone, the glottal correlation (GCORR) algorithm takes advantage of non-acoustic throat sensors such as the general electromagnetic motion sensor (GEMS). The non–acoustic sensor provides a measure of the glottal excitation function that is relatively immune to background acoustic noise. Thus, inspired by human speech production mechanisms, the GCORR algorithm extracts the desired speech signal from noisy acoustic mixture using statistical correlation between the speech and its excitation. The algorithm leads to a significant reduction of wide–band noise, even when the SNR is very low. The improvement in the quality of the speech is demonstrated in terms of an objective evaluation.

Summary form only given. Field-programmable analog arrays (FPAAs) provide a method for rapidly prototyping analog systems. Currently available commercial and academic FPAAs are typically based on operational amplifiers (or other similar analog primitives) with only a few computational elements per chip. While their specific architectures vary, their small sizes and often restrictive interconnect designs leave current FPAAs limited in functionality, flexibility, and usefulness. We explore the use of floating-gate devices as the core programmable element in a large-scale FPAA with applications in signal processing emphasized. An FPAA architecture is presented that offers increased functionality and flexibility in realizing analog signal processing systems, and experimental data from a testbed FPAA is shown. In addition, mainstream signal processing systems are discussed that can be effectively implemented on large-scale reconfigurable analog devices thereby realizing dramatic savings in power over traditional digital solutions and improved time-to-market over traditional analog designs.

A novel channel-optimized multistage vector quantization (COMSVQ) codec is presented in which the stage codebooks are jointly designed. The proposed codec uses a signal source and channel-dependent distortion measure to encode line spectral frequencies derived from segments of a speech signal. Simulation results are provided to demonstrate the consistent reduction in the spectral distortion obtained using the proposed codec as compared to the conventional sequentially designed channel-matched multistage vector quantizer.

The direct use of vector quantization (VQ) to encode LPC parameters in a communication system suffers from the following two limitations: 1) complexity of implementation for large vector dimensions and codebook sizes and 2) sensitivity to errors in the received indices due to noise in the communication channel. In the past, these issues have been simultaneously addressed by designing channel matched multistage vector quantizers (CM-MSVQ). A sub-optimal sequential design procedure has been used to train the codebooks of the CM-MSVQ. In this paper, a novel channel-optimized multistage vector quantization (CO-MSVQ) codec is presented, in which the stage codebooks are jointly designed. The proposed codec uses a source and channel-dependent distortion measure to encode line spectral frequencies derived from segments of a speech signal. Extensive simulation results are provided to demonstrate the consistent reduction in both the mean and the variance of the spectral distortion obtained using the proposed codec relative to the conventional sequentially designed CM-MSVQ. Furthermore, the perceptual quality of the reconstructed speech using the proposed codec was found to be better than that obtained using the sequentially designed CM-MSVQ.

Removal of ambient noise from a single-channel audio signal is becoming an increasingly important problem due to the proliferation of portable communication devices. Furthermore, in applications such as wireless telephony and phonetic data mining, it is desired that noise suppression be robust to changing noise conditions and that processing take place in real time or faster. This paper proposes an adaptive noise suppression system that mitigates or eliminates processing artifacts common to Wiener filtering without decreasing speech recognition performance. Results of one implementation of such a structure demonstrate significant improvements in both perceptual quality and speech recognition performance under noisy conditions.

We introduce the concept of cooperative analog-digital signal processing, and its implications on real-time signal processing functions. We discuss some of the trade-offs between performing operations in the analog computing circuit and digital computing circuits, and a general framework for “cooperative analog/digital signal processing systems” is presented. A key aspect of the cooperative analog/digital signal processing systems is the fact that new advances in analog VLSI circuits make it possible to developed advanced analog signal processing systems with programmable elements.

In this paper, we propose a continuous-time audio noise suppression algorithm for reducing stationary background noise in a single microphone signal. The algorithm is targeted for implementation in sub-threshold analog floating-gate transistor circuits for extremely low-power systems. The noise suppression algorithm divides the noisy signal into exponentially spaced sub-bands and estimates the noisy signal envelope and the noise envelope to calculate a time-varying gain for each sub-band based on each band's a posteriori SNR. A sigmoidal gain function in for use in the analog system is introduced and a normalized SNR is also proposed for consistent noise suppression for different input noise variances. Simulations of the system show promising results.

The recent rapid increase in applications of DSP has generated a need for technical personnel with a DSP background. Until recently, most DSP education has been at the graduate level, so many practicing engineers are developing a need for additional DSP education. At Georgia Institute of Technology, we are delivering an online continuing education course that is targeted at these engineers. The 12-week course is comprised of three interconnected tracks in DSP system theory, real-time implementation principles, and laboratory assignments with MATLAB and the latest fixed-point TI DSP chips. The course was designed after a needs analysis performed by Texas Instruments training specialists. It is intended specifically to provide a full semester's course in a way that is adapted to the schedules and circumstances of practicing engineers. We describe both the content of the course and the technology for delivering it online.

The sinusoidal transform (ST) provides a sparse representation for speech signals by utilizing several psychoacoustic phenomena. It is well suited to applications in signal enhancement because the signal is represented in a parametric manner that is easy to manipulate. The multi-resolution sinusoidal transform (MRST) has the additional advantage that it is both particularly well suited to typical speech signals and well matched to the human auditory system. The currently reported work discusses the removal of noise from a noisy signal by applying an adaptive Wiener filter to the MRST parameters and then conditioning the parameters to eliminate "musical noise". In informal tests MRST based noise reduction was found to reduce background noise significantly better than traditional Wiener filtering and to virtually eliminate the "musical noise" often associated with Wiener filtering.

The sinusoidal transform, as developed by Quatieri and McAulay (1986), provides a sparse representation for speech signals by taking advantage of psychoacoustic masking. The currently reported work takes the sinusoidal transform one step further by considering the frequency resolution abilities of the human auditory system in more detail. The new transform is based on the wavelet principle of variable resolution in time/frequency analysis. Specifically, a sinusoidal transform is developed which uses quadrature mirror filter (QMF) banks to obtain better time resolution at high frequencies and better frequency resolution at low frequencies. This naturally provides a perceptually improved allocation of the sinusoids. The new transform matches the human auditory system better than its predecessor and it also matches speech signals well, both fricative sounds and voiced speech. The QMF based ST is then shown to be equivalent to a more efficient FFT based implementation.

A new hearing compensation algorithm based on a homomorphic multiplicative AGC (automatic gain control) is evaluated and compared against commercially available digitally programmable analog hearing aids. Both quantitative (speech recognition threshold and speech discrimination) and qualitative tests (estimation of perceived quality) were used in the evaluation. The new algorithm is shown to have made significant progress in restoring normal or near normal hearing for hearing impaired individuals.

A model is proposed which mathematically transforms an acoustic stimulus into a form which is believed to be more nearly related to that used by the auditory cortex to interpret the sound. The model is based upon research towards understanding the response of the human auditory system to sound stimuli. The motivation and approach in developing this model follow the philosophy pursued in the development of a similar model to the problem of hearing compensation for impaired individuals is shown to yield a bank of bandpass filters each followed by a homomorphic multiplicative AGC. Clinical tests on hearing impaired subjects suggest that this approach is far superior to other current hearing compensation schemes.