Additive Background Noise Research Articles

Noise compensation for speech recognition with arbitrary additive noise

This paper investigates speech recognition involving additive background noise, assuming no knowledge about the noise characteristics. A new method, namely universal compensation (UC), is proposed as a solution to the problem. The UC method is an extension of the missing-feature method, i.e., recognition based only on reliable data but robust to any corruption type, including full corruption in which the noise affects all time-frequency components of the speech representation. The UC technique achieves robustness to unknown, full noise corruption through a novel combination of the multicondition training method and the missing-feature method. Multicondition training is employed to convert fullband spectral corruption into partial-band spectral corruption, which is achieved by training the model using data involving simulated wide-band noise at different signal-to-noise ratios. The missing-feature principle is employed to reduce the effect of the remaining partial-band corruption on recognition by basing the recognition only on the matched or compensated spectral components from the multicondition training. The combination of these two strategies makes the new method potentially capable of dealing with arbitrary additive noise-with arbitrary temporal-spectral characteristics-based only on clean speech training data and simulated noise data, without requiring knowledge of the actual noise. Two databases, Aurora 2 and an E-set word database, have been used to evaluate the UC method. Experiments on Aurora 2 indicate that the new model has the potential to achieve a recognition performance close to the performance obtained by a multicondition baseline model trained using data involving the test environments. Further experiments for noise conditions unseen in Aurora 2 show significant performance improvement for the new model over the multicondition model. The experimental results on the E-set database demonstrate the ability of the UC model to deal with acoustically confusing recognition tasks.

Robust speech recognition in noisy environments based on subband spectral centroid histograms

We investigate how dominant-frequency information can be used in speech feature extraction to increase the robustness of automatic speech recognition against additive background noise. First, we review several earlier proposed auditory-based feature extraction methods and argue that the use of dominant-frequency information might be one of the major reasons for their improved noise robustness. Furthermore, we propose a new feature extraction method, which combines subband power information with dominant subband frequency information in a simple and computationally efficient way. The proposed features are shown to be considerably more robust against additive background noise than standard mel-frequency cepstrum coefficients on two different recognition tasks. The performance improvement increased as we moved from a small-vocabulary isolated-word task to a medium-vocabulary continuous-speech task, where the proposed features also outperformed a computationally expensive auditory-based method. The greatest improvement was obtained for noise types characterized by a relatively flat spectral density.

Toward Robust Speech Recognition and Understanding

The principal cause of speech recognition errors is a mismatch between trained acoustic/language models and input speech due to the limited amount of training data in comparison with the vast variation of speech. It is crucial to establish methods that are robust against voice variation due to individuality, the physical and psychological condition of the speaker, telephone sets, microphones, network characteristics, additive background noise, speaking styles, and other aspects. This paper overviews robust architecture and modeling techniques for speech recognition and understanding. The topics include acoustic and language modeling for spontaneous speech recognition, unsupervised adaptation of acoustic and language models, robust architecture for spoken dialogue systems, multi-modal speech recognition, and speech summarization. This paper also discusses the most important research problems to be solved in order to achieve ultimate robust speech recognition and understanding systems.

Additive background noise as a source of non-linear mismatch in the cepstral and log-energy domain

The aim of this investigation is to determine to what extent automatic speech recognition may be enhanced if, in addition to the linear compensation accomplished by mean and variance normalisation, a non-linear mismatch reduction technique is applied to the cepstral and energy features, respectively. An additional goal is to determine whether the degree of mismatch between the feature distributions of the training and test data that is associated with acoustic mismatch, differs for the cepstral and energy features. Towards these aims, two non-linear mismatch reduction techniques – time domain noise reduction and histogram normalisation – were evaluated on the Aurora2 digit recognition task as well as on a continuous speech recognition task with noisy test conditions similar to those in the Aurora2 experiments. The experimental results show that recognition performance is enhanced by the application of both non-linear mismatch reduction techniques. The best results are obtained when the two techniques are applied simultaneously. The results also reveal that the mismatch in the energy features is quantitatively and qualitatively much larger than the corresponding mismatch associated with the cepstral coefficients. The most substantial gains in average recognition rate are therefore accomplished by reducing training-test mismatch for the energy features.

Noise compensation for speech recognition with arbitrary additive noise

A new method for noise compensation for additive background noise based only on clean speech training data is described, assuming arbitrary noise characteristics. Experiments on Aurora 2 indicate that the new method has achieved a performance comparable to, or better than, the performance obtained by the baseline model trained on multi-condition data.

Noise adaptive speech recognition based on sequential noise parameter estimation

In this paper, a noise adaptive speech recognition approach is proposed for recognizing speech which is corrupted by additive non-stationary background noise. The approach sequentially estimates noise parameters, through which a non-linear parametric function adapts mean vectors of acoustic models. In the estimation process, posterior probability of state sequence given observation sequence and the previously estimated noise parameter sequence is approximated by the normalized joint likelihood of active partial paths and observation sequence given the previously estimated noise parameter sequence. The Viterbi process provides the normalized joint-likelihood. The acoustic models are not required to be trained from clean speech and they can be trained from noisy speech. The approach can be applied to perform continuous speech recognition in presence of non-stationary noise. Experiments conducted on speech contaminated by simulated and real non-stationary noise show that when acoustic models are trained from clean speech, the noise adaptive speech recognition system provides improvements in word accuracy as compared to the normal noise compensation system (which assumes the noise to be stationary) in slowly time-varying noise. When the acoustic models are trained from noisy speech, the noise adaptive speech recognition system is found to be helpful to get improved performance in slowly time-varying noise over a system employing multi-conditional training.

Optimum nonlinear composite filter for distortion-tolerant pattern recognition.

We describe a nonlinear distortion-tolerant filter for pattern recognition that is optimum in terms of tolerance to input noise and discrimination capability. This filter was derived by minimization of the output energy that is due to the overlapping additive noise and the input scene, and the output of the filter meets the design constraints obtained from the training data set. The performance of this filter was tested with an input scene containing one of the training data sets, a nontraining true target, and a false object in the presence of overlapping additive noise and nonoverlapping background noise. We carried out Monte Carlo runs to measure the statistical performance of the filter and obtained receiver operating characteristics curves to show the detection capabilities of the filter.

Total Internal Reflection Microscopy: Distortion Caused by Additive Noise

Total internal reflection microscopy (TIRM) is an optical technique for monitoring Brownian fluctuations in the separation between a single microscopic sphere and a flat plate in aqueous media. The sphere is levitated above the plate by colloidal forces such as double-layer or steric repulsion. Changes in elevation as small as 1 nm can be detected by measuring the light scattered by a single sphere when illuminated by an evanescent wave. From the Boltzmann distribution of elevations sampled by the sphere over a long time, the potential energy (PE) profile can be determined with a resolution of about 0.1kT. By corrupting clean data (having a mean scattering intensity of Īs and a standard deviation of σs) obtained by Brownian dynamics simulations with various levels of additive background noise (having a mean background intensity of Īb and a standard deviation of σb), we show how the PE profiles obtained from TIRM are distorted. Increasing Īb narrows the profile and shifts it toward smaller elevations; conv...

Compensation of channel and noise distortions combining normalization and speech enhancement techniques

This paper introduces two techniques to obtain robust speech recognition devices in mismatch conditions (additive noise mismatch and channel mismatch). The first algorithm, adaptive Gaussian attenuation algorithm (AGA), is a speech enhancement technique developed to reduce the effects of additive background noise in a wide range of signal noise ratio (SNR) and noise conditions. The technique is closely related to the classical noise spectral subtraction (SS) scheme, but in the proposed model the mean and variance of noise are used to better attenuate the noise. Information of the SNR is also introduced to provide adaptability at different SNR conditions. The second algorithm, cepstral mean normalization and variance-scaling technique (CMNVS), is an extension of the cepstral mean normalization (CMN) technique to provide robust features to convolutive and additive noise distortions. The requirements of the techniques are also analyzed in the paper. Combining both techniques the relative channel distortion effects were reduced to 90% on the HTIMIT task and the relative additive noise effects were reduced to 77% using the TIMIT database mixed with car noises at different SNR conditions.

On the temporal decorrelation of feature parameters for noise-robust speech recognition

We propose a new frame decorrelation method for robust speech recognition in noisy environments. In most cases, signal perturbation is caused by channel distortion and additive background noise, and can be modeled as a slowly varying term in either the log spectral or the linear-spectral domains. Thus, it is effective to deemphasize slowly varying stationary components in the spectral feature domain of speech signals, which can be considered as a temporal decorrelation process. The proposed method presents a well structured high-pass filter using the decorrelation principle, and provides some significant insights into existing high-pass approaches, such as relative spectral (RASTA) processing. The performance of the proposed method was evaluated by speaker-independent isolated-word recognition experiments using the hidden Markov model (HMM). Noisy speech was simulated by adding noise sources taken from the Noisex-92 database. Experimental results showed that the proposed method was effective for speech recognition with significant noise and yielded better performance than other high-pass methods. In addition, we compared the dynamic property of the proposed filter with that of delta features. The feature obtained by the proposed method may offer most of the delta feature property.

Improved restoration of space object imagery

We examine methods for preprocessing a collection of atmospheric turbulence-degraded short-exposure imagery to improve the resolving power of estimation algorithms. We redefine the method known as frame selection in the context of optimizing estimation results. We compare several measures of image quality with idealized standards, demonstrating their relative ability to rank highly the least-degraded image frames. In particular, we find the Fisher information measure to be the most noise tolerant and robust frame-selection measure. We then examine the resolving implication of removing additive background noise resulting from the sky and telescope. Specifically, we show that background compensation acts as a de facto restoration of the compact object support and leads to furthering the resolving power of estimation methods. Results from simulated imaging scenarios demonstrate the improved ability of a multiframe maximum a posteriori estimator to restore the passband object distribution as well as to further recover the lost spectral content residing beyond the diffraction limit.

Evaluation of PEMO in robust speech recognition

A major problem in speech recognition is the robustness of recognition performance against additive background noise and convolutive distortions due to changing transmission channels or varying microphone characteristics. One approach to overcome this problem is to apply PEMO [Dau, Püschel, and Kohlrausch, J. Acoust. Soc. Am. 99, 3615–3622 (1996)] for extraction of noise robust feature vectors. The recognition performance achievable with PEMO in combination with a locally recurrent neural network (LRNN) or hidden Markov models (HMM) for feature scoring was evaluated in a task of speaker-independent word recognition. The robustness was tested and compared with that of other feature types by applying the speech recognition systems to speech signals disturbed by additive background noise and to speech signals recorded over telephone lines. It was found that only with LRNN it is possible to exploit the potential of PEMO, not with HMM. A detailed analysis of the interplay between PEMO features and the scoring techniques revealed that the long-term dependencies in the sequence PEMO features can only be captured by LRNN. Furthermore, LRNN discriminate between the distinct and sparse peaks of the PEMO speech representation—which are well maintained also in the case of noisy speech—and artifacts introduced by the additive or convolutive noise.

Generalized optimum receiver for pattern recognition with multiplicative, additive, and nonoverlapping background noise

We design an optimum receiver to detect a pattern or a reference signal. We design a receiver that detects the signal distorted by a multiplicative noise on the signal itself, as well as by additive noise and by nonoverlapping scene noise. We design the optimum receiver under the condition in which the statistics of the multiplicative and nonoverlapping scene noise are not available. In the case in which additive noise is present and the statistics of the multiplicative noise are not known, the usual method of maximizing the likelihood function to estimate the statistics of stationary noise fails. We overcome this problem by viewing the noise processes as vector random variables and describe two different schemes to estimate the statistics of the multiplicative noise. Using computer simulations we show that, for the images tested here, the optimum receiver performs better than some of the existing receivers.

ICARUS: Source generator based real-time recognition of speech in noisy stressful and Lombard effect environments ☆

The problem of real-time automatic speech recognition in an adverse environment is addressed. Though much research has been performed in the area of speech recognition, only limited success has been demonstrated for real-time recognition in noisy stressful environments. The primary reason for this is that the performance of present day recognition algorithms are predicated on the assumptions of the environmental settings in which the algorithms have been formulated and implemented. In this paper, we discuss the effects of additive background noise on speech quality and recognition parameters, and propose a source generator based framework to address stress and noise. Using this framework, a computationally efficient real-time recognition system called ICARUS is developed. The speech recognition system incorporates direct processing steps to address the effects of additive noise on the speech signal and stress on the speech production system. Central issues which are addressed include (i) improved characterization of speech spoken in noisy situations involving both parameter estimation methods and analysis of varying speech characteristics spoken in adverse environments (i.e., stress and Lombard effect), (ii) exploration of signal processing strategies tailored to such speech, and (iii) demonstration of real-time system performance of the proposed methods. The proposed recognition system was formulated using a digital signal processing platform. Performance evaluations showed an improvement in speech feature representation under stressed speaking conditions, with an average improvement in recognition rate of +17.28% across eleven noisy stressful speaking conditions.

Objective speech quality assessment and the RPE-LTP coding algorithm in different noise and language conditions

The formulation of reliable signal processing algorithms for speech coding and synthesis require the selection of a prior criterion of performance. Though coding efficiency (bits/second) or computational requirements can be used, a final performance measure must always include speech quality. In this paper, three objective speech quality measures are considered with respect to quality assessment for American English, noisy American English, and noise-free versions of seven languages. The purpose is to determine whether objective quality measures can be used to quantify changes in quality for a given voice coding method, with a known subjective performance level, as background noise or language conditions are changed. The speech coding algorithm chosen is regular-pulse excitation with long-term prediction (RPE-LTP), which has been chosen as the standard voice compression algorithm for the European Digital Mobile Radio system. Three areas are considered for objective quality assessment which include: (i) vocoder performance for American English in a noise-free environment, (ii) speech quality variation for three additive background noise sources, and (iii) noise-free performance for seven languages which include English, Japanese, Finnish, German, Hindi, Spanish, and French. It is suggested that although existing objective quality measures will never replace subjective testing, they can be a useful means of assessing changes in performance, identifying areas for improvement in algorithm design, and augmenting subjective quality tests for voice coding/compression algorithms in noise-free, noisy, and/or non-English applications.

Performance of a correlation demodulator in the presence of a speech signal

The performance of a correlation demodulator (CD) in a channel with an additive speech signal and background noise is analyzed. Expression of eenergy loss due to the influence of the speech signal and of probability of error are obtained. For that purpose the speech signal is considered as a random combination of vowel, consonant and a pause. Moreover, the vowels as well as the consonants are modelled as response of some linear time invariant (LTI) systems to white noise. As an illustration the case of an antipodal (BPSK) signal is considered. The theoretical results are in very good agreement with experiment.

The internal signal-to-noise ratio in hearing-impaired listeners

The reduced frequency resolution often demonstrated by listeners with sensorineural hearing impairment might have at least two related effects on their perception of speech in background noise. The internal representation of the peaks and valleys in the speech spectrum become smeared, resulting in less precise frequency analysis than can be achieved by normal-hearing persons. In addition, when the speech is embedded in noise, poorer frequency resolution will, in theory, result in a greater-than-normal decrease in the signal-to-noise ratio at the outputs of the auditory filters. This poorer-than-normal internal S/N might account for the common observation that noise is more degrading to speech understanding by hearing-impaired listeners than by normal-hearing listeners. To assess the validity of this notion, both normal-hearing and hearing-impaired subjects participated in an identification experiment using highly simplified vowel-like harmonic complexes presented in quiet and in three levels of broadband noise. After equating individual performance in terms of percent correct identification in quiet, trading relationships were established between amount of spectral contrast in the stimuli (level differences between peaks and valleys in the spectra) and levels of the background noise. When the effect of spectral smearing introduced by reduced frequency analysis was compensated by increasing the input spectral contrast, the effects of additional background noise were similar for normal-hearing and hearing-impaired groups of subjects. Results from individual listeners will be discussed in light of the degree of impaired frequency resolution as determined by auditory filter shape measurements. [Work supported by NIH.]

Local-geometric-projection method for noise reduction in chaotic maps and flows.

We describe a method for noise reduction in chaotic systems that is based on projection of the set of points comprising an embedded noisy orbit in ${\mathit{openR}}^{\mathit{d}}$ toward a finite patchwork of best-fit local approximations to an m-dimensional surface M'\ensuremath{\subset}${\mathit{openR}}^{\mathit{d}}$, m\ensuremath{\le}d. We generate the orbits by the delay coordinate construction of Ruelle and Takens [N. H. Packard et al., Phys. Rev. Lett. 45, 712 (1980); F. Takens, in Dynamical Systems and Turbulence, Warwick, 1980, edited by D. A. Rand and L.-S. Young (Springer, Berlin, 1981)] from time series v(t), which in an experimental situation we would assume to have come, together with additional high-dimensional background noise, from an underlying dynamical system ${\mathit{f}}^{\mathit{t}}$: M\ensuremath{\rightarrow}M existing on some low m-dimensional manifold M. The surface M' in ${\mathit{openR}}^{\mathit{d}}$ is the assumed embedded image of M. We give results of systematic studies of linear (tangent plane) projection schemes. We describe in detail the basic algorithm for implementing these schemes. We apply the algorithm iteratively to known map and flow time series to which white noise has been added. In controlled studies, we measure the signal-to-noise ratio improvements, iterating ${\mathit{n}}_{\mathit{M}}$ times until a stable maximum ${\mathrm{\ensuremath{\delta}}}_{\mathit{M}}$ is achieved. We present extensive results for ${\mathrm{\ensuremath{\delta}}}_{\mathit{M}}$ and ${\mathit{n}}_{\mathit{M}}$ for a wide range of values of embedding trial dimension d, projection dimension k, number of nearest-neighbor points for local approximation \ensuremath{\nu}, embedding delay \ensuremath{\Delta}, sampling interval \ensuremath{\Delta}T, initial noise amplitude scrN, and trajectory length N.We give results for very low and very high noise amplitudes 0%\ensuremath{\le}scrN\ensuremath{\le}100%. We develop an empirical method for estimating the initial noise level for a given experimental time series, and for the optimal choice of algorithm parameters to achieve peak reduction. We present interesting results of application of the noise-reduction algorithm to a chaotic time series produced from a periodically driven magnetoelastic ribbon experiment on the control of chaos. Two noteworthy elements of the noise-reduction method we describe result in certain stabilizing and efficiency features. The first is our use of a physical replacement time series, which is a unique scalar time series with the property that its corresponding delay coordinate construction data state vector time series in ${\mathit{openR}}^{\mathit{d}}$ is optimally close to the noise-reduced replacement vector time series generated by the projection. The second is the introduction of a ``measure-ordered'' cover, which produces notable improvement in reliability, control, and computational efficiency of the whole algorithm.

Chaos, multiplicity, crisis, and synchronicity in higher-order neural networks.

We study a randomly diluted higher-order network of spinlike neurons that interact via Hebbian-type connections and derive and solve exact dynamical equations for a general block-sequential updating algorithm. The system has a variety of static and oscillatory solutions. The bifurcation parameters in the present model include neuronal interaction coefficients, the synchronicity parameter, and a rescaled noise level, which represents the combined effects of the random synaptic dilution, interference between stored patterns, and additional background noise.

Monitoring the efficiency of hearing aid fitting in the aged by the social hearing handicap index

The unsatisfactory results obtained from tone and speech audiometric studies in identifying the effects of hearing disability, in particular the psychosocial factors, prompted us to adopt a psychological test to determine the "social hearing ability". The test consists of 20 questions which all relate to "social hearing ability" in urban society. Ten questions concern hearing ability in situations where background noise plays either a negligible role or none at all. The remaining 10 questions concern selectivity ability, that is hearing with additional background noise. The evaluation of the questionnaire results in a classification of the so-called Social Hearing Handicap Index (SHHI) in percentages together with a breakdown of the deafness components and the selectivity components into percentages. This differentiation provides additional information about the demands placed on the hearing ability of humans in quiet or noisy surroundings. The information obtained from the test enables longitudinal studies on the course of a hearing disability or the effectivity of a particular therapy, e.g. hearing aid, to be carried out in the elderly.