Log Power Spectrum Research Articles

Identification of Language using Mel-Frequency Cepstral Coefficients (MFCC)

This paper focuses on the task of identifying a language from speech signal. In this paper, we have use Mel-frequency cepstral coefficient as features. Language identification models are developed for fifteen Indian languages namely Assamese, Bangla, Guajarati, Hindi, Kannada, Kashmiri, Malayalam, Marathi, Nepali, Oriya, Punjabi, Rajasthani, Tamil, Telugu and Urdu using these spectral features. The identification of above mentioned languages is carried out using Gaussian mixture model. A Semi natural read database is used for obtaining the language specific information. MFCC is obtained by using linear cosine transform of log power spectrum on a nonlinear mel-frequency scale. This paper shows that the performance of Language identification system is better when trained and tested with twenty nine features as compared to six, eight, thirteen, nineteen and twenty one MFCC features. It means more the number of features we use better the result we get. The average language recognition rate over fifteen Indian languages is around 88\\%.

Crustal thermal regime of Ikogosi warm spring, Nigeria inferred from aeromagnetic data

Spectral analysis method was applied to aeromagnetic data obtained for Ikogosi warm spring (IWS) area of southwestern Nigeria. This was done with the objective of determining the bottom of the magnetized crust called Curie point depth (CDP) and understand the nature and extent of the local geothermal system at depth beneath IWS. The depth to the centroid, Zo, of the deepest distribution of the magnetic dipoles was obtained by computing least-squares fit to the lowest-frequency segment of the azimuthally averaged log power spectrum. The average depth to the top of the deepest crustal block was computed as the depth to the top, Zt, of the second lowest-frequency segment of the spectrum. The depth to the bottom of the deepest magnetic dipoles, the inferred Curie point depth, was then calculated from Zb = 2Zo − Zt. The Curie depth estimates for IWS range between 4.68 and 11.38 km (below sea level). We also estimate the heat flow and Curie temperature using a one-dimensional conductive heat transport model. The average heat flow, 42 mW m−2, and geothermal gradient, 32°C/km, obtained suggest a low enthalpy thermal regime. The Curie temperature for the region varies between 153°C and 350°C. Also, an inverse linear relationship between heat flow and Curie depths was determined. Good agreement between the Curie point depths derived from heat flow data and magnetic data suggests that the Curie point depth analysis is useful to estimate the regional thermal structure and the tectonic settings.

Vertical torque responses to vestibular stimulation in standing humans

The effects of electrical vestibular stimulation upon movement and perception suggest two evoked sensations: head roll and inter-aural linear acceleration. The head roll vector causes walking subjects to turn in a direction dependent on head pitch, requiring generation of torque around a vertical axis. Here the effect of vestibular stimulation upon vertical torque (T(z)) was investigated during quiet stance. With the head tilted forward, square-wave stimuli applied to the mastoid processes evoked a polarity-specific T(z) response accompanied by trunk yaw. Stochastic vestibular stimulation (SVS) was used to investigate the effect of head pitch with greater precision; the SVS–T(z) cross-correlation displayed a modulation pattern consistent with the head roll vector and this was also reflected by changes in coherence at 2–3 Hz. However, a separate response at 7–8 Hz was unaffected by head pitch. Head translation (rather than rotation) had no effect upon this high frequency response either, suggesting it is not caused by a sense of body rotation induced by an inter-aural acceleration vector offset from the body. Instead, high coherence between medio-lateral shear force and T(z) at the same frequency range suggests it is caused by mechanical coupling to evoked medio-lateral sway. Consistent with this explanation, the 7–8 Hz response was attenuated by 90 deg head roll or yaw, both of which uncouple the inter-aural axis from the medio-lateral sway axis. These results demonstrate two vertical torque responses to electrical vestibular stimulation in standing subjects. The high frequency response can be attributed to mechanical coupling to evoked medio-lateral sway. The low frequency response is consistent with a reaction to a sensation of head roll, and provides a novel method for investigating proprioceptive-vestibular interactions during stance.

Research on Ground Penetrating Radar (GPR) Exploration on Subgrade Failure Based on Spectrum Profile

Subgrade failure is relative to spectrum signal received by ground penetrating radar. For the practical exploring demand of Beijing road surface, the method to identify hidden diseases in subgrade such as cavity and loose zone is studied. On the basis of summarizing the layer response to electromagnetic wave and analyzing the limitation identifying deterioration cause, the viewpoint of transforming time section into spectrum section for explanation is put forward. The spectrum response of electromagnetic wave from cavity is of high sensitivity, so spectrum section obtained by using the method of log power spectrum and gauss window function improves the identifying deterioration capability. The testing result shows that spectrum technology is efficient to analyze hidden deterioration causes in subgrade.

EEG-Based Lapse Detection With High Temporal Resolution

A warning system capable of reliably detecting lapses in responsiveness (lapses) has the potential to prevent many fatal accidents. We have developed a system capable of detecting lapses in real-time with second-scale temporal resolution. Data was from 15 subjects performing a visuomotor tracking task for two 1-hour sessions with concurrent electroencephalogram (EEG) and facial video recordings. The detector uses a neural network with normalized EEG log-power spectrum inputs from two bipolar EEG derivations, though we also considered a multichannel detector. Lapses, identified using a combination of video rating and tracking behavior, were used to train our detector. We compared detectors employing tapped delay-line linear perceptron, tapped delay-line multilayer perceptron (TDL-MLP), and long short-term memory (LSTM) recurrent neural networks operating continuously at 1 Hz. Using estimates of EEG log-power spectra from up to 4 s prior to a lapse improved detection compared with only using the most recent estimate. We report the first application of a LSTM to an EEG analysis problem. LSTM performance was equivalent to the best TDL-MLP network but did not require an input buffer. Overall performance was satisfactory with area under the curve from receiver operating characteristic analysis of 0.84 +/- 0.02 (mean +/- SE) and area under the precision-recall curve of 0.41 +/- 0.08.

Comparative Study of Speaker Identification Methods: dPLRM, SVM and GMM

A comparison of performances is made of three text-independent speaker identification methods based on dual Penalized Logistic Regression Machine (dPLRM), Support Vector Machine (SVM) and Gaussian Mixture Model (GMM) with experiments by 10 male speakers. The methods are compared for the speech data which were collected over the period of 13 months in 6 utterance-sessions of which the earlier 3 sessions were for obtaining training data of 12 seconds' utterances. Comparisons are made with the Mel-frequency cepstrum (MFC) data versus the log-power spectrum data and also with training data in a single session versus in plural ones. It is shown that dPLRM with the log-power spectrum data is competitive with SVM and GMM methods with MFC data, when trained for the combined data collected in the earlier three sessions. dPLRM outperforms GMM method especially as the amount of training data becomes smaller. Some of these findings have been already reported in [1]--[3].

Reference-free quantification of EEG spectra: Combining current source density (CSD) and frequency principal components analysis (fPCA)

Objective Definition of appropriate frequency bands and choice of recording reference limit the interpretability of quantitative EEG, which may be further compromised by distorted topographies or inverted hemispheric asymmetries when employing conventional (non-linear) power spectra. In contrast, fPCA factors conform to the spectral structure of empirical data, and a surface Laplacian (2-dimensional CSD) simplifies topographies by minimizing volume-conducted activity. Conciseness and interpretability of EEG and CSD fPCA solutions were compared for three common scaling methods. Methods Resting EEG and CSD (30 channels, nose reference, eyes open/closed) from 51 healthy and 93 clinically-depressed adults were simplified as power, log power, and amplitude spectra, and summarized using unrestricted, Varimax-rotated, covariance-based fPCA. Results Multiple alpha factors were separable from artifact and reproducible across subgroups. Power spectra produced numerous, sharply-defined factors emphasizing low frequencies. Log power spectra produced fewer, broader factors emphasizing high frequencies. Solutions for amplitude spectra showed optimal intermediate tuning, particularly when derived from CSD rather than EEG spectra. These solutions were topographically distinct, detecting multiple posterior alpha generators but excluding the dorsal surface of the frontal lobes. Instead a low alpha/theta factor showed a secondary topography along the frontal midline. Conclusions CSD amplitude spectrum fPCA solutions provide simpler, reference-independent measures that more directly reflect neuronal activity. Significance A new quantitative EEG approach affording spectral components is developed that closely parallels the concept of an ERP component in the temporal domain.

Revised gravity for offshore India and the isostatic compensation of submarine features

High-resolution free air gravity data have been derived for offshore India covering the area 60°E to 104°E and 5°S to 26°N using reprocessed radar altimeter data from the geodetic missions of ERS-1 and GEOSAT. The free air gravity gives enhanced imaging of geological features such as the Murray, Laxmi, Chagos-Laccadive, 85°E and 90°E ridges. These data provide a major improvement compared with the existing satellite-derived dataset of Sandwell and Smith v. 9.2. This improvement has been quantified by determining the standard deviation of the difference grid between these two satellite-derived gravity datasets which is ± 10 mGal for the whole study area, with a variation of ± 5 mGal over the active ridges. Plots of log power spectrum versus wavenumber of these datasets and correlation coefficients between these datasets and marine data confirm the high resolution of the data presented. The local (Airy) and regional (flexure) model of isostasy are tested at 9 locations covering a range of different geological features. The northern and southern parts of the eastern continental margins of India (blocks 3 and 4) show different flexural responses suggesting they conform to extension and transform type margins, respectively, whereas the western continental margin and Laxmi ridge (blocks 8 and 9) have flexural responses conforming more to a fragmented margin. The deeper oceanic features of the 90°E and Chagos-Laccadive ridges have distinctly different isostatic responses conforming to flexure and Airy type responses, respectively. This suggests the origin of the 90°E ridge formed on top of a pre-existing crust (like a hot spot trace), whereas the Chagos-Laccadive ridge formed close to an active ridge spreading centre.

Curie Point Depths Based on Spectrum Analysis of Aeromagnetic Data, West Anatolian Extensional Province, Turkey

The Curie point depth map of Western Anatolia was constituted from spectral analysis of the aeromagnetic data. The Curie point depth values from 53 overlapping blocks, 90 × 90 km in size, have been estimated from the band-pass filtered data. The slope of the longest wavelength part of the radially averaged log power spectrum divided by the radial frequency produced the depth to the centroid (z0) for the deepest crustal block. The depth to the top (zt) was obtained by the slope of the second longest wavelength part of the spectrum. From these depths, the Curie point depth was then calculated by using zb=2z0−zt. The Curie point depth estimates for Western Anatolia range between 8.2 and 19.9 km.

Curie Point Depths of Macedonia and Thrace, N. Greece

— The aeromagnetic data of Macedonia and Thrace were used to produce Curie point estimates. The data were high pass filtered to remove components arising from topography and magnetic core fields which were not adequately modeled by a DGRF. The depth to the centroid, z 0, of the deepest distribution of the magnetic dipoles was obtained by computing a least-squares fit to the lowest-frequency segment of the azimuthally averaged log power spectrum. The average depth to the top of the deepest crustal block was computed as the depth to the top, z t , of the second lowest-frequency segment of the spectrum. The depth to the bottom of the deepest magnetic dipoles, the inferred Curie point depth, was then calculated from z b =2z 0 − z t . The Curie depth estimates for Macedonia and Thrace range between 11.2 and 17.3 km. These results are consistent with the depths inferred by extrapolating known geothermal gradient and heat-flow values.

Method and apparatus for extracting speech related facial features for use in speech recognition systems

The apparatus for the recognition of speech comprises an acoustic preprocessor, a visual preprocessor, and a speech classifier that operates the acoustic and visual preprocessed data. The acoustic preprocessor comprises a log mel spectrum analyzer that produces an equal mel bandwidth log power spectrum. The visual processor detects the motion of a set of fiducial markers on the speaker's face and extracts a set of normalized distance vectors describing lip and mouth movement. The speech classifier uses a multilevel time-delay neural network operating on the preprocessed acoustic and visual data to form an output probability distribution that indicates the probability of each candidate utterance having been spoken, based on the acoustic and visual data.

Analysis of potential field anomalies in Lavrion mining area, Greece

Mining activities in Lavrion began during the first millennium B.C. after the decline of ancient Athens and then restarted more deliberately during the nineteenth century. Aeromagnetic data from a 1967 survey of the mining area was recompiled, processed, and interpreted for the present study. The original flight lines were digitized and leveled, and the international geomagnetic reference field (IGRF) was removed. The data were inverted by means of a terracing technique that defines separate domains of uniform distribution of physical properties that cause the magnetic anomalies. The log power spectrum was computed; along with the results of terracing, it suggested the existence of two sources of the magnetic anomaly. The long‐wavelength anomaly reflects a large, concealed body that is most probably a granitic intrusion, consistent with local geological evidence. The source of the short‐wavelength anomaly is a strongly magnetized body attributed to the net effect of various thin, magnetite‐bearing sulfide zones. The anomalies were then separated in the wavenumber domain. Magnetic susceptibility measurements were made in situ on the exposed parts of the local formations. Three‐dimensional models whose effect simulates the observed anomalies were calculated. Results of the modeling show that the large magnetic body is buried at 0.68 km depth. The small, relatively shallow body is about 0.035 km thick and buried at 0.6 km depth. The bodies do not show any corresponding gravity anomaly on the regional Bouguer gravity anomaly map.

Curie Point Depth of the Island of Crete (Greece)

—The aeromagnetic data of the island of Crete were inverted to produce Curie point estimates. The data were high-pass filtered to remove components arising from topography and magnetic core fields which were not adequately modeled by IGRF. The depth to the centroid, z 0 , of the deepest distribution of the magnetic dipoles, was obtained by computing a least-squares fit to the lowest-fre quency segment of the azimuthally averaged log power spectrum. The depth to the top of the deepest crustal block was computed as the depth, z t , to the centroid of the second deepest distribution, using the second lowest-frequency segment of the spectrum. The depth to the bottom of the deepest dipoles, the inferred Curie point depth, is then z b = 2z 0−z t . The Curie depth estimates range between 24 and 28 km. This is in accordance with the depths inferred by extrapolating heat-flow values measured in boreholes.

Neural network acoustic and visual speech recognition system

The apparatus for the recognition of speech comprises an acoustic preprocessor, a visual preprocessor, and a speech classifier that operates on the acoustic and visual preprocessed data. The acoustic preprocessor comprises a log mel spectrum analyzer that produces an equal mel bandwidth log power spectrum. The visual processor detects the motion of a set of fiducial markers on the speaker's face and extracts a set of normalized distance vectors describing lip and mouth movement. The speech classifier uses a multilevel time-delay neural network operating on the preprocessed acoustic and visual data to form an output probability distribution that indicates the probability of each candidate utterance having been spoken, based on the acoustic and visual data.

Neural network acoustic and visual speech recognition system training method and apparatus

The apparatus for the recognition of speech includes an acoustic preprocessor, a visual preprocessor, and a speech classifier that operates on the acoustic and visual preprocessed data. The acoustic preprocessor comprises a log mel spectrum analyzer that produces an equal mel bandwidth log power spectrum. The visual processor detects the motion of a set of fiducial markers on the speaker's face and extracts a set of normalized distance vectors describing lip and mouth movement. The speech classifier uses a multilevel time-delay neural network operating on the preprocessed acoustic and visual data to form an output probability distribution that indicates the probability of each candidate utterance having been spoken, based on the acoustic and visual data. The training system includes the speech recognition apparatus and a control processor with an associated memory. Noisy acoustic input training data together with visual data is used to generate acoustic and visual feature training vectors for processing by the speech classifier. A control computer adjusts the synaptic weights of the speech classifier based upon the noisy input training data and exemplar output vectors for producing a robustly trained classifier based on the analogous visual counterpart of the Lombard effect.

On the cause of 1/f‐power spectral scaling in borehole sonic logs

Borehole logs provide in situ information about the scaling of fluctuations in the elastic properties of rock. A surprising finding has been the widespread 1/fβ‐type (0.5 < β < 1.5) scaling of borehole sonic log power spectra, yielding no clear correlation with geological age, type or history, raising questions about the reliability and origin of these findings. Ubiquitous fractures, for example, have been evoked as the most likely explanation of these results. Firstly, we show that the smoothing effects of the logging tool must be taken into account in order to obtain reliable values of β. We then investigate the scaling properties of lithology only in perfectly exposed metamorphic rock, by explicitly excluding fractures. We find clear 1/f scaling for reconstructed velocity data based solely on lithological variations. Although fractures are certainly important in contributing to 1/f fluctuations in borehole logs, intrinsic geology may also play a significant role.

A set of windows for high resolution spectral analysis

A set of windows whose normalized half main-lobe widths fall between those of rectangular and the Hanning or Hamming windows, is proposed. These windows are compared with Kaiser's near-optimum zeroth-order Bessel and first-order Bessel windows. The results show that the new windows compare very favourably with Kaiser's windows. The distinct advantage of the windows is their very simple form, similar to that of a Hamming window. As an example of the use of these windows, an estimate of the log power spectrum of a signal consisting of two sinusoidal tones of unequal amplitudes is considered and compared with the estimates using other windows.

Separation Filtering Applied to Aeromagnetic Data

Separation filtering of regional and residual gravity and magnetic fields has been the subject of extensive research. Wavenumber and spatial domain methods, variously described as matched filtering, layer filtering, depth slices etc. have been developedSeparation or layer filtering of magnetic data allows us to deconvolve the effects of causative sources occurring around a particular level. In practice, it is impossible to achieve a complete separation since the problem is non-linear. Separation filtering depends fundamentally on the concept of random distributions of sources within discrete layers and assumes that there is no statistical difference in response along each ideal layer and no correlation between the distribution in each layer.Interpretation of the log-power spectrum provides slope and intercept parameters for filter design. It is necessary to be able to recognise distinct linear segments of the spectrum corresponding to well separated depth ensembles in the data in order to deconvolve the effects of each depth ensemble with minimum contamination from sources at different levels. Non-linear spectra cannot be interpreted. It is important to apply a finite source width correction to the raw spectrum, otherwise indicated source depths will be too deep.Tests have been carried out on spectral matched filtering, differential upward continuation separation filtering, exponential taper filtering and a modified second vertical derivative filter. All filtering was performed in wavenumber domain. Spectral analysis of the test area suggested three depth ensembles with average depths of 1500 m, 300 m and 20 m below surface. The modified second vertical derivative technique provided the best representation of the shallow source ensemble being more stable and with a larger dynamic range than the matched filter residual filter. The spectral matched filter regional provided the best representation of the deep ensemble, although there was some contamination by shallower sources. The intermediate ensemble could not be extracted with any degree of confidence.Separation filtering can make a useful contribution to the enhancement of magnetic data, provided the limitations in the methods are recognised. Even where spatial source separation is reasonable, the filter will not manage to perform a complete rejection of shallow sources or deeper sources and will distort intermediate sources even more. The results should be used qualitatively to assist in source discrimination rather than used as input to analytical techniques.

Presence of cepstral peak in random reflected ultrasound signals.

A peak in the power cepstrum, or the squared magnitude of the Fourier transform of the data log-power spectrum, is commonly observed when processing reflections from plate-like structures, such as membranes. In this case, the cepstral peak at the smallest nonzero time lag, or quefrency, value can be used to determine the thickness of the plate. For reflections from a medium composed of randomly distributed scatterers, such as liver tissue, a cepstral peak is also commonly observed, but cannot be so intuitively explained as in the deterministic case above. In this paper, it is demonstrated that the presence of a cepstral peak depends on the form of the probability density function (pdf) of the separation between reflectors. In the case where the pdf is uniform from O to SM, the cepstral peak is found to occur at the quefrency corresponding to SM. For simple unimodal pdfs, a cepstral peak will occur at the location of the maximum probability. These observations are shown analytically and verified through simulations. The diagnostic value of these results lies in the interpretation of the relation of the cepstral peak location to the spacing of the scattering elements in the tissue.

Presence of cepstral peak in random reflected ultrasound signals

A peak in the power cepstrum, or the squared magnitude of the Fourier transform of the data log-power spectrum, is commonly observed when processing reflections from plate-like structures, such as membranes. In this case, the cepstral peak at the smallest nonzero time lag, or quefrency, value can be used to determine the thickness of the plate. For reflections from a medium composed of randomly distributed scatterers, such as liver tissue, a cepstral peak is also commonly observed, but cannot be so intuitively explained as in the deterministic case above. In this paper, it is demonstrated that the presence of a cepstral peak depends on the form of the probability density function (pdf) of the separation between reflectors. In the case where the pdf is uniform from 0 to S M, the cepstral peak is found to occur at the quefrency corresponding to S M. For simple unimodal pdfs, a cepstral peak will occur at the location of the maximum probability. These observations are shown analytically and verified through simulations. The diagnostic value of these results lies in the interpretation of the relation of the cepstral peak location to the spacing of the scattering elements in the tissue.