Effect Of Residual Noise Research Articles

An investigation of neural uncertainty estimation for target speaker extraction equipped RNN transducer

Target-speaker speech recognition aims to recognize the speech of an enrolled speaker from an environment with background noise and interfering speakers. This study presents a joint framework that combines time-domain target speaker extraction and recurrent neural network transducer (RNN-T) for speech recognition. To alleviate the adverse effects of residual noise and artifacts introduced by the target speaker extraction module to the speech recognition back-end, we explore to training the target speaker extraction and RNN-T jointly. We find a multi-stage training strategy that pre-trains and fine-tunes each module before joint training is crucial in stabilizing the training process. In addition, we propose a novel neural uncertainty estimation that leverages useful information from the target speaker extraction module to further improve the back-end speech recognizer (i.e., speaker identity uncertainty and speech enhancement uncertainty). Compared to a recognizer with target speech extraction front-end, our experiments show that joint-training and the neural uncertainty module reduce 7% and 17% relative character error rate (CER) on multi-talker simulation data, respectively. The multi-condition experiments indicate that our method can reduce 9% relative CER in the noisy condition without losing performance in the clean condition. We also observe consistent improvements in further evaluation of real-world data based on vehicular speech.

Multichannel Statistical Broadband Wavelet Deconvolution for Improving Resolution of Seismic Signals

Many popular deconvolution methods based on Robinson's convolutional model have played an important role in improving the temporal resolution of seismic data. However, the outcomes of applying these deconvolution methods to real land seismic data are not always desirable due to the effect of noise in the deconvolution process. Although the noise in the seismogram can be minimized during the recording process, the effect of residual noise on deconvolution operators can result in poor deconvolution output. To address the shortcomings of conventional deconvolution methods, we developed a new deconvolution method based on a multichannel statistical principle. In the proposed method, we have extended the surface-consistent convolutional model to include a noise component, thus including the noise effect on deconvolution operators in the deconvolution process. According to the proposed multichannel statistical strategy, we first calculated the autocorrelation of the seismogram, in which the lateral variation effect on the wavelet is considered because of inhomogeneities in the vicinity of sources and receivers. Then, we adopted a local fitting technique to approximate the autocorrelation of the seismic wavelet. To obtain the seismic data with a broad bandwidth and low-noise level, we used the integral-Ricker wavelet as the desired output wavelet. Tests on synthetic data and real land seismic data demonstrate the effectiveness of the proposed method in increasing the resolution of seismic signals.

Noise Reduction Method of Underwater Acoustic Signals Based on Uniform Phase Empirical Mode Decomposition, Amplitude-Aware Permutation Entropy, and Pearson Correlation Coefficient.

Noise reduction of underwater acoustic signals is of great significance in the fields of military and ocean exploration. Based on the adaptive decomposition characteristic of uniform phase empirical mode decomposition (UPEMD), a noise reduction method for underwater acoustic signals is proposed, which combines amplitude-aware permutation entropy (AAPE) and Pearson correlation coefficient (PCC). UPEMD is a recently proposed improved empirical mode decomposition (EMD) algorithm that alleviates the mode splitting and residual noise effects of EMD. AAPE is a tool to quantify the information content of nonlinear time series. Unlike permutation entropy (PE), AAPE can reflect the amplitude information on time series. Firstly, the original signal is decomposed into a series of intrinsic mode functions (IMFs) by UPEMD. The AAPE of each IMF is calculated. The modes are separated into high-frequency IMFs and low-frequency IMFs, and all low-frequency IMFs are determined as useful IMFs (UIMFs). Then, the PCC between the high-frequency IMF with the smallest AAPE and the original signal is calculated. If PCC is greater than the threshold, the IMF is also determined as a UIMF. Finally, all UIMFs are reconstructed and the denoised signal is obtained. Chaotic signals with different signal-to-noise ratios (SNRs) are used for denoising experiments. Compared with EMD and extreme-point symmetric mode decomposition (ESMD), the proposed method has higher SNR and smaller root mean square error (RMSE). The proposed method is applied to noise reduction of real underwater acoustic signals. The results show that the method can further eliminate noise and the chaotic attractors are smoother and clearer.

Uniform Phase Empirical Mode Decomposition: An Optimal Hybridization of Masking Signal and Ensemble Approaches.

The empirical mode decomposition (EMD) is an established method for the time-frequency analysis of nonlinear and nonstationary signals. However, one major drawback of the EMD is the mode mixing effect. Many modifications have been made to resolve the mode mixing effect. In particular, disturbance-assisted EMDs, such as the noise-assisted EMD and the masking EMD, have been proposed to resolve this problem. These disturbance-assisted approaches have led to a better performance of the EMD in the analysis of real-world data sets, but they may also have two side effects: the mode splitting and residual noise effects. To minimize or eliminate the mode mixing effect while avoiding the two side effects of traditional disturbance-assisted EMDs, we propose an EMD-based algorithm assisted by sinusoidal functions with a designed uniform phase distribution with a comprehensive theoretical explanation for the substantial reduction of the mode splitting and the residual noise effects simultaneously. We examine the performance of the new method and compare it to those of other disturbance-assisted EMDs using synthetic signals. Finally numerical experiments with real-world examples are conducted to verify the performance of the proposed method.

Robust first quantization matrix estimation based on filtering of recompression artifacts for non-aligned double compressed JPEG images

This paper presents a novel method for first quantization matrix estimation for non-aligned double JPEG compressed images. The non-aligned double JPEG scenario poses several challenges in the form of recompression artifacts which needs to be handled effectively in order to arrive at accurate first quantization estimate. One such recompression artifact which is prominent in the non-aligned scenario is the blocking artifacts arising due to the misalignment of the DCT grids in the successive compression cycles. The proposed method investigates the impact of the blocking artifacts induced errors on the DCT histogram and counters these errors via a novel DCT histogram filtering strategy. In addition, the residual noise which is also a recompression artifact is countered utilizing the local rank transform which adaptively filters the residual noise effects. The filtered histograms are compared with a synthetically created ideal second quantization matrix to estimate the degree of similarity between the two cases. Utilizing the maximum degree of similarity, the selection of first quantization value is performed. Experimental analysis utilizing several non-aligned double compressed JPEG images taken from UCID and RAISE datasets with Quality Factor (QF) of second compression being greater than the first compression (QF2>QF1) along with the comparative analysis with one of the state-of-the-art method shows accurate estimation of the first quantization values for the proposed method. The proposed method finds its application in image forensics as well as in steganalysis.

Refinement of Enhanced Speech Using Hybrid-Median Filter and Harmonic Regeneration Approach

This study proposes a post-processor to improve the harmonic structure of a vowel in an enhanced speech, enabling the speech quality to be improved. Initially, a speech enhancement algorithm is employed to reduce the background noise for a noisy speech. Hence the enhanced speech is post-processed by a hybrid-median filter to reduce the musical effect of residual noise. Since the harmonic spectra are impacted by background noise and a speech enhancement process, the quality of a vowel is deteriorated. A harmonic regenerated method is developed to improve the quality of post-processed speech. Experimental results show that the proposed method can improve the quality of post-processed speech by adequately regenerating harmonic spectra.

Reduction of musical residual noise using block-and-directional-median filter adapted by harmonic properties

Many speech enhancement systems can efficiently remove background noise. However, most of them suffer from musical residual noise which is very annoying to the human ear. This study proposes a post-processing system to efficiently reduce the effect of musical residual noise, enabling the enhanced speech to be improved. Noisy speech is firstly enhanced by a speech enhancement algorithm to reduce background noise. The enhanced speech is then post-processed by a block-and-directional-median (BDM) filter adapted by harmonic properties, causing the musical effect of residual noise being efficiently reduced. In the case of a speech-like spectrum, directional-median filtering is performed to slightly reduce the musical effect of residual noise, where a strong harmonic spectrum of a vowel is well maintained. The quality of post-processed speech is then ensured. On the contrary, block-median filtering is performed to greatly reduce the spectral variation in noise-dominant regions, enabling the spectral peaks of musical tones to be significantly smoothed. The musical effect of residual noise is therefore reduced. Finally, the preprocessed and post-processed spectra are integrated according to the speech-presence probability. Experimental results show that the proposed post processor can efficiently improve the performance of a speech enhancement system by reducing the musical effect of residual noise.

Evaluation of the potential of MODIS satellite data to predict vegetation phenology in different biomes: An investigation using ground-based NDVI measurements

Vegetation phenology is the study of the timing of seasonal events that are considered to be the result of adaptive responses to climate variations on short and long time scales. In the field of remote sensing of vegetation phenology, phenological metrics are derived from time series of optical data. For that purpose, considerable effort has been specifically focused on developing noise reduction and cloud-contaminated data removal techniques to improve the quality of remotely-sensed time series. Comparative studies between time series composed of satellite data acquired under clear and cloudy conditions and from radiometric data obtained with high accuracy from ground-based measurements constitute a direct and effective way to assess the operational use and limitations of remote sensing for predicting the main plant phenological events. In the present paper, we sought to explicitly evaluate the potential use of MODerate resolution Imaging Spectroradiometer (MODIS) remote sensing data for monitoring the seasonal dynamics of different types of vegetation cover that are representative of the major terrestrial biomes, including temperate deciduous forests, evergreen forests, African savannah, and crops. After cloud screening and filtering, we compared the temporal patterns and phenological metrics derived from in situ NDVI time series and from MODIS daily and 16-composite products. We also evaluated the effects of residual noise and the influence of data gaps in MODIS NDVI time series on the identification of the most relevant metrics for vegetation phenology monitoring. The results show that the inflexion points of a model fitted to a MODIS NDVI time series allow accurate estimates of the onset of greenness in the spring and the onset of yellowing in the autumn in deciduous forests (RMSE≤one week). Phenological metrics identical to those provided with the MODIS Global Vegetation Phenology product (MDC12Q2) are less robust to data gaps, and they can be subject to large biases of approximately two weeks or more during the autumn phenological transitions. In the evergreen forests, in situ NDVI time series describe the phenology with high fidelity despite small temporal changes in the canopy foliage. However, MODIS is unable to provide consistent phenological patterns. In crops and savannah, MODIS NDVI time series reproduce the general temporal patterns of phenology, but significant discrepancies appear between MODIS and ground-based NDVI time series during very localized periods of time depending on the weather conditions and spatial heterogeneity within the MODIS pixel. In the rainforest, the temporal pattern exhibited by a MODIS 16-day composite NDVI time series is more likely due to a pattern of noise in the NDVI data structure according to both rainy and dry seasons rather than to phenological changes. More investigations are needed, but in all cases, this result leads us to conclude that MODIS time series in tropical rainforests should be interpreted with great caution.

Objective threshold estimation and measurement of the residual background noise in auditory evoked potentials of goldfish

A survey of papers using auditory evoked potentials (AEPs) published over the last 10 years (Table I) demonstrates that most AEP studies in animals have used subjective methods for auditory threshold determination. Subjective methods greatly reduce the value of statistical hypothesis testing and jeopardize tests of hypothetical experimental group differences in hearing sensitivity. Correspondingly, many attempts have been made to develop objective threshold determination methods, but these have not been used widely. Further, they seldom include an appreciation of the effects of residual noise in the AEP. In this study, AEPs evoked by tonal and noise stimuli in goldfish (Carassius auratus) were recorded and the residual background noise was measured and analyzed in detail. High variability was found in residual noise, but can be effectively controlled with a simple modification of averaging routines. Considerable interobserver disagreements were found using subjective threshold estimation. An objective method of threshold determination was developed based on comparison between AEP amplitude and controlled residual noise, using a signal detection theory approach to set specific threshold criteria. The usefulness of AEP in hypothesis testing for auditory function requires more control over residual background noise amplitudes and the use of objective threshold determination techniques.

Backscattering measurements of atmospheric aerosols at CO_2 laser wavelengths: implications of aerosol spectral structure on differential-absorption lidar retrievals of molecular species

The volume backscattering coefficients of atmospheric aerosol were measured with a tunable CO2 lidar system at various wavelengths in Utah (a desert environment) along a horizontal path a few meters above the ground. In deducing the aerosol backscattering, a deconvolution (to remove the smearing effect of the long CO2 lidar pulse and the lidar limited bandwidth) and a constrained-slope method were employed. The spectral shape beta(lambda) was similar for all the 13 measurements during a 3-day period. A mean aerosol backscattering-wavelength dependence beta(lambda) was computed from the measurements and used to estimate the error Delta(CL) (concentration-path-length product) in differential-absorption lidar measurements for various gases caused by the systematic aerosol differential backscattering and the error that is due to fluctuations in the aerosol backscattering. The water-vapor concentration-path-length product CL and the average concentration C = <CL>/L for a path length L computed from the range-resolved lidar measurements is consistently in good agreement with the water-vapor concentration measured by a meteorological station. However, I was unable to deduce, reliably, the range-resolved water-vapor concentration C(r), which is the derivative of the range-dependent product CL, because of the effect of residual noise caused mainly by errors in the deconvolved lidar measurements.

Spectra of auditory brainstem responses and spontaneous EEG.

Human auditory brainstem responses (ABR's) are sensory evoked potentials that can be recorded within a few milliseconds following a transient acoustic stimulus. This paper describes results of a simulation study that evaluated alternative techniques for estimating the spectrum of the ABR signal and of the background EEG. The effects of residual noise in the average signal, the number of responses in the average, and the use of smoothing windows were considered. Spectra of human ABR's were also obtained. The spectral energy decreases with frequency to 2 kHz and seems to flatten above 2 kHz. The ABR signal spectrum is limited to below approximately 1.5 kHz. There are three main components: a low-frequency component around 100 Hz, a midfrequency component around 500 Hz, and a high-frequency component at 1000 Hz. Narrow-band filtering suggests that the midfrequency component is primarily associated with Jewett wave V, while the high-frequency component is primarily associated with the early Jewett waves. These two components are probably the most important for determining the latencies of waves. The low-frequency component appears to correspond to a slow wave on which the Jewett waves are superimposed.