Effects of microphone type on acoustic measures of voice.

Acoustic measures provide an objective means to describe pathological voices and are a routine component of the clinical voice examination. Because the voice sample is obtained using a microphone, microphone characteristics have the potential to influence the values of parameters obtained from a voice sample. This project examined how the choice of microphone affects key voice parameters and investigated how one might compensate for such microphone effects through filtering or by including additional parameters in the decision process. A database of 53 normal voice samples and 100 pathological voice samples was used in four experiments conducted in an anechoic chamber using four different microphones. One omnidirectional microphone and three cardioid microphones were used in these experiments. The original voice samples were presented to each microphone through a speaker located in an anechoic chamber, and the output of each microphone sampled to computer disk. Each microphone modified the frequency spectrum of the voice signal; this, in turn, affected the values of the voice parameters obtained. These microphone effects reduced the accuracy with which acoustic measures of voice could be used to discriminate pathological from normal voices. Discrimination performance improved when the microphone output was filtered to compensate for microphone frequency response. Performance also improved when spectral moment coefficient parameters were added to the vocal function parameters already in use.

Acoustic discrimination of pathological voice: sustained vowels versus continuous speech.

We investigated the ability of acoustic measures to discriminate between normal and pathological talkers. Two groups of measures were compared: (a) those extracted from sustained vowels and (b) those based on continuous speech samples. Nine acoustic measures, which include fundamental frequency and amplitude perturbation measures, long term average spectral measures, and glottal noise measures were extracted from both sustained vowel and continuous speech samples. Our experiments were performed on a published database of 53 normal talkers and 175 talkers with a pathological voice. The classification performance of the nine acoustic measures was quantified using linear discriminant analysis and receiver operating characteristic (ROC) curve analysis. When individual measures were considered in isolation, classification was more accurate for measures extracted from sustained vowels than for those based on continuous speech samples. Classification accuracy improved when combinations of acoustic parameters were considered. For such combinations of measures, classification results were comparable for measures extracted from continuous speech samples and for those based on sustained vowels.

Identification of pathological voices using glottal noise measures.

We investigated the abilities of four fundamental frequency (F0)-dependent and two F0-independent measures to quantify vocal noise. Two of the F0-dependent measures were computed in the time domain, and two were computed using spectral information from the vowel. The F0-independent measures were based on the linear prediction (LP) modeling of vowel samples. Tests using a database of sustained vowel samples, collected from 53 normal and 175 pathological talkers, showed that measures based on the LP model were much superior to the other measures. A classification rate of 96.5% was achieved by a parameter that quantifies the spectral flatness of the unmodeled component of the vowel sample.

A comparison of high precision F0 extraction algorithms for sustained vowels.

Perturbation analysis of sustained vowel waveforms is used routinely in the clinical evaluation of pathological voices and in monitoring patient progress during treatment. Accurate estimation of voice fundamental frequency (F0) is essential for accurate perturbation analysis. Several algorithms have been proposed for fundamental frequency extraction. To be appropriate for clinical use, a key consideration is that an F0 extraction algorithm be robust to such extraneous factors as the presence of noise and modulations in voice frequency and amplitude that are commonly associated with the voice pathologies under study. This work examines the performance of seven F0 algorithms, based on the average magnitude difference function (AMDF), the input autocorrelation function (AC), the autocorrelation function of the center-clipped signal (ACC), the autocorrelation function of the inverse filtered signal (IFAC), the signal cepstrum (CEP), the Harmonic Product Spectrum (HPS) of the signal, and the waveform matching function (WM) respectively. These algorithms were evaluated using sustained vowel samples collected from normal and pathological subjects. The effect of background noise and of frequency and amplitude modulations on these algorithms was also investigated, using synthetic vowel waveforms.

Convergence characteristics of two algorithms in non-linear stimulus artefact cancellation for electrically evoked potential enhancement.

Somatosensory evoked potentials (SEPs) are a sub-class of evoked potentials (EPs) that are very useful in diagnosing various neuromuscular disorders and in spinal cord and peripheral-nerve monitoring. Most often, the measurements of these signals are contaminated by stimulus-evoked artefact. Conventional stimulus-artifact (SA) reduction schemes are primarily hardware-based and rely on some form of input blanking during the SA phase. This procedure can result in partial SEP loss if the tail of the SA interferes with the SEP. Adaptive filters offer an attractive solution to this problem by iteratively reducing the SA waveform while leaving the SEP intact. Owing to the inherent non-linearities in the SA generation system, non-linear adaptive filters (NAFs) are most suitable. SA reduction using NAFs based on truncated second-order Volterra expansion series is investigated. The focus is on the performance of two main adaptation algorithms, the least mean square (LMS) and recursive least squares (RLS) algorithms, in the context of non-linear adaptive filtering. A comparison between the convergence and performance characteristics of these two algorithms is made by processing both simulated and experimental SA data. It is found that, in high artefact-to-noise ratio (ANR) SA cancellation, owing to the large eigenvalue spreads, the RLS-based NAF is more efficient than the LMS-based NAF. However, in low-ANR scenarios, the RLS- and LMS-based NAFs exhibit similar convergence properties, and the computational simplicity of the LMS-based NAFs makes them the preferred option.

Adaptive stimulus artifact reduction in noncortical somatosensory evoked potential studies.

Somatosensory evoked potentials (SEP's) are an important class of bioelectric signals which contain clinically valuable information. The surface measurements of these potentials are often contaminated by a stimulus evoked artifact. The stimulus artifact (SA), depending upon the stimulator and measurement system characteristics, may obscure some of the information carried by the SEP's. Conventional methods for SA reduction employ hardware-based circuits which attempt to eliminate the SA by blanking the input during SA period. However, there is a danger of losing some of the important SEP information, especially if the stimulating and recording electrodes are close together. In this paper, we apply both linear and nonlinear adaptive filtering techniques to the problem of SA reduction. Nonlinear adaptive filters (NAF's) based on truncated second-order Volterra series expansion are discussed and their applicability to SA cancellation is explored through processing both simulated and in vivo SEP data. The performances of the NAF and the finite impulse response (FIR) linear adaptive filter (LAF) are compared by processing experimental SEP data collected from different recording sites. Due to the inherent nonlinearities in the generation of the SA, the NAF is shown to achieve significantly better SA cancellation compared to the LAF.

Multireference adaptive noise cancellation applied to somatosensory evoked potentials.

Somatosensory Evoked Potentials (SEP's) contain information that is useful in diagnosing various physiological disorders. However, surface measurements of these potentials suffer from very poor Signal-to-Noise ratio (SNR) resulting in imperceptible SEP waveforms. This factor motivates the employment of dedicated signal processing techniques to improve the quality of the waveform. The objective of this research work is to improve the SNR of SEP by eliminating the predominant myoelectric interference. The strategy followed to achieve this goal is to process the SEP signal by MultiReference Adaptive Noise Cancellation (MRANC). A theoretical model for the MRANC is presented and its performance under the influence of various factors is investigated and compared with other signal processing techniques. The performance of the MRANC is then evaluated by processing simulated and in vivo SEP data. It is found that the MRANC gives a significant improvement in the SNR of the SEP.