Panos E. Papamichalis

Hands-free voice communication in an automobile with a microphone array

The authors present the result of their research on developing a hands-free voice communication system with a microphone array for use in an automobile environment. The goal of this research is to develop a speech acquisition and enhancement system so that a speech recognizer can reliably be used inside a noise automobile environment, for digital cellular phone application. Speech data have been collected using a microphone array and a digital audio tape (DAT) recorder inside a real car for several idling and driving conditions, and processed using delay-and-sum and adaptive beamforming algorithms. Performance criteria including signal-to-noise ratio and speech recognition error rate have been evaluated for the processed data. Detailed performance results presented show that the microphone array is superior to a single microphone.<<ETX>>

The TMS320C30 floating-point digital signal processor

The 320C30 is a fast processor with a large memory space and floating-point-arithmetic capabilities. The authors describe the 320C30 architecture in detail, discussing both the internal organization of the device and the external interfaces. They also explain the pipeline structure, addressing software-related issues and constructs, and examine the development tools and support. Finally, they present examples of applications. Some of the major features of the 320C30 are: a 60-ns cycle time that results in execution of over 16 million instructions per second (MIPS) and over 33 million floating-point operations per second (Mflops); 32-bit data buses and 24-bit address buses for a 16M-word overall memory space; dual-access, 4 K*32-bit on-chip ROM and 2 K*32-bit on-chip RAM; a 64*32-bit program cache; a 32-bit integer/40-bit floating-point multiplier and ALU; eight extended-precision registers, eight auxiliary registers, and 23 control and status registers; generally single-cycle instructions; integer, floating-point, and logical operation; two- and three-operand instructions; an on-chip DMA controller; and fabrication in 1- mu m CMOS technology and packaging in a 180-pin package. These facilitate FIR (finite impulse response) and IIR (infinite impulse response) filtering, telecommunications and speech applications, and graphics and image processing applications.<<ETX>>

Speech analysis/synthesis system with energy normalization and silence suppression

Energy normalization in speech synthesis systems is achieved by a look-ahead adaptive normalization procedure, wherein energy is adaptively tracked, and the adaptive energy-tracking value is used to normalize a much earlier frames energy value.

Time encoding of LPC roots

A time encoding scheme of the center frequencies of the LPC inverse filter roots is presented. The roots are represented in terms of center frequency (CF) and bandwidth (BW), and the continuity of the CFs in time is established by a dynamic programming scheme which has a cost function depending on both CFs and BWs. Segmentation points are set at the beginning or end of root tracks, at the voiced-unvoiced transitions and at the peaks of a function measuring the dissimilarity of adjacent frames. The tracks within segments are then fitted by linear combinations of orthogonal polynomials. Without quantization, the processed speech is subjectively indistinguishable from the original synthetic speech. Good quality is achieved even below 1000 bps.

Robust color image superresolution: an adaptive M-estimation framework

This paper introduces a new color image superresolution algorithm in an adaptive, robust M-estimation framework. Using a robust error norm in the objective function, and adapting the estimation process to each of the low-resolution frames, the proposed method effectively suppresses the outliers due to violations of the assumed observation model, and results in color superresolution estimates with crisp details and no color artifacts, without the use of regularization. Experiments on both synthetic and real sequences demonstrate the superior performance over using the and error norms in the objective function.

Graph-Laplacian Features for Neural Waveform Classification

Analysis of extracellular recordings of neural action potentials (known as spikes) is highly dependent upon the accuracy of neural waveform classification, commonly referred to as spike sorting. Feature extraction is an important stage of this process because it can limit the quality of clustering that is performed in the feature space. Principal components analysis (PCA) is the most commonly used feature extraction method employed for neural spike recordings. To improve upon PCAs feature extraction performance for neural spike sorting, we revisit the PCA procedure to analyze its weaknesses and describe an improved feature extraction method. This paper proposes a linear feature extraction technique that we call graph-Laplacian features, which simultaneously minimizes the graph Laplacian and maximizes variance. The algorithms performance is compared with PCA and a wavelet-coefficient-based feature extraction algorithm on simulated single-electrode neural data. A cluster-quality metric is proposed to quantitatively measure the algorithm performance. The results show that the proposed algorithm produces more compact and well-separated clusters compared to the other approaches.

FFT implementation on the TMS320C30

The implementation of several FFT (fast Fourier transform) algorithms on the TMS320C30, the third-generation device in the Texas Instruments family of digital signal processors is reported. The algorithms considered are the complex radix-2 and radix-4, and real-valued radix-2 FFT. The architecture and the instruction set of the TMS320C30 permit flexible and compact coding of the algorithms in assembly language while preserving close correspondence to a high-level language implementation. The efficiency of the architecture and the speed of the device (60 ns) make possible the realization of a 1024-point FFT in 3.75 ms (complex, radix-2), 3.04 ms (complex, radix-4), or 1.67 ms (real, radix-2).<<ETX>>

A graph-laplacian-based feature extraction algorithm for neural spike sorting

Analysis of extracellular neural spike recordings is highly dependent upon the accuracy of neural waveform classification, commonly referred to as spike sorting. Feature extraction is an important stage of this process because it can limit the quality of clustering which is performed in the feature space. This paper proposes a new feature extraction method (which we call Graph Laplacian Features, GLF) based on minimizing the graph Laplacian and maximizing the weighted variance. The algorithm is compared with Principal Components Analysis (PCA, the most commonly-used feature extraction method) using simulated neural data. The results show that the proposed algorithm produces more compact and well-separated clusters compared to PCA. As an added benefit, tentative cluster centers are output which can be used to initialize a subsequent clustering stage.

An adaptive M-estimation framework for robust image super resolution without regularization

This paper introduces a new image super-resolution algorithm in an adaptive, robust M-estimation framework. Super-resolution reconstruction is formulated as an optimization (minimization) problem whose objective function is based on a robust error norm. The effectiveness of the proposed scheme lies in the selection of a specific class of robust M-estimators, redescending M-estimators, and the incorporation of a similarity measure to adapt the estimation process to each of the low-resolution frames. Such a choice helps in dealing with violations to the assumed imaging model that could have generated the low-resolution frames from the unknown high-resolution one. The proposed approach effectively suppresses the outliers without the use of regularization in the objective function, and results in high-resolution images with crisp details and no artifacts. Experiments on both synthetic and real sequences demonstrate the superior performance over methods based on the L2 and L1 in the objective function.

Graph-spectrum-based neural spike features for stereotrodes and tetrodes

Extracellular recording of neural signals records the action potentials (known as spikes) of neurons adjacent to the electrode as well as the noise generated by the overall neural activity around the electrode. Analysis of these spikes is highly dependent upon the accuracy of neural waveform classification, commonly referred to as spike sorting. Feature extraction is an important stage of this process because it can limit the quality of clustering which is performed in the feature space. This paper introduces a new feature extraction algorithm for neural spike sorting to isolate single neuronal units out of multi-unit activity when multiple closely-spaced electrodes (two for a stereotrode, four for a tetrode) are used. The proposed algorithm, which is inspired by spectral graph theory, simultaneously minimizes the graph-Laplacian and maximizes the variance. Real test signals from stereotrode and tetrode recordings show that the proposed approach outperforms the most commonly-used feature extraction methods, including Principal Components Analysis (PCA) and ratios of peak spike amplitudes between different electrodes of a stereotrode or tetrode.