Several Signal-to-noise Ratios Research Articles

Increasing the Robustness of i-vectors with Model Compensated First Order Statistics

Speaker recognition systems achieved significant improvements over the last decade, especially due to the performance of the i-vectors. Despite the achievements, mismatch between training and test data affects the recognition performance considerably. In this paper, a solution is offered to increase robustness against additive noises by inserting model compensation techniques within the i-vector extraction scheme. For stationary noises, the model compensation techniques produce highly robust systems. Parallel Model Compensation and Vector Taylor Series are considered as state-of-the-art model compensation techniques. Applying these methods to the first order statistics, a noisy total variability space training is aimed, which will reduce the mismatch resulted by additive noises. All other parts of the conventional i-vector scheme remain unchanged, such as total variability matrix training, reducing the i-vector dimensionality, scoring the i-vectors. The proposed method was tested with four different noise types with several signal to noise ratios (SNR) from -6 dB to 18 dB with 6 dB steps. High reductions in equal error rates were achieved with both methods, even at the lowest SNR levels. On average, the proposed approach produced more than 50% relative reduction in equal error rate.

Development and Evaluation of a Language-Independent Test of Auditory Discrimination for Referrals for Cochlear Implant Candidacy Assessment.

Objectives:The purpose of this study was to (1) develop a Language-independent Test of Auditory Discrimination (LIT-AD) between speech sounds so that people with hearing loss who derive limited speech perception benefits from hearing aids (HAs) may be identified for consideration of cochlear implantation and (2) examine the relationship between the scores for the new discrimination test and those of a standard sentence test for adults wearing either HAs or cochlear implants (CIs).Design:The test measures the ability of the listener to correctly discriminate pairs of nonsense syllables, presented as sequential triplets in an odd-one-out format, implemented as a game-based software tool for self-administration using a tablet computer. Stage 1 included first a review of phonemic inventories in the 40 most common languages in the world to select the consonants and vowels. Second, discrimination testing of 50 users of CIs at several signal to noise ratios (SNRs) was carried out to generate psychometric functions. These were used to calculate the corrections in SNR for each consonant-pair and vowel combination required to equalize difficulty across items. Third, all items were individually equalized in difficulty and the overall difficulty set. Stage 2 involved the validation of the LIT-AD in English-speaking listeners by comparing discrimination scores with performance in a standard sentence test. Forty-one users of HAs and 40 users of CIs were assessed. Correlation analyses were conducted to examine test–retest reliability and the relationship between performance in the two tests. Multiple regression analyses were used to examine the relationship between demographic characteristics and performance in the LIT-AD. The scores of the CI users were used to estimate the probability of superior performance with CIs for a non-CI user having a given LIT-AD score and duration of hearing loss.Results:The LIT-AD comprises 81 pairs of vowel–consonant–vowel syllables that were equalized in difficulty to discriminate. The test can be self-administered on a tablet computer, and it takes about 10 min to complete. The software automatically scores the responses and gives an overall score and a list of confusable items as output. There was good test–retest reliability. On average, higher LIT-AD discrimination scores were associated with better sentence perception for users of HAs (r = −0.54, p <0.001) and users of CIs (r = −0.73, p <0.001). The probability of superior performance with CIs for a certain LIT-AD score was estimated, after allowing for the effect of duration of hearing loss.Conclusions:The LIT-AD could increase access to CIs by screening for those who obtain limited benefits from HAs to facilitate timely referrals for CI candidacy evaluation. The test results can be used to provide patients and professionals with practical information about the probability of potential benefits for speech perception from cochlear implantation. The test will need to be evaluated for speakers of languages other than English to facilitate adoption in different countries.

Speech discrimination impairments as a marker of disease severity in multiple sclerosis

BackgroundMultiple Sclerosis (MS) pathology is likely to disrupt central auditory pathways, thereby affecting an individual's ability to discriminate speech from noise. Despite the importance of speech discrimination in daily communication, it's characterization in the context of MS remains limited. This cross-sectional study evaluated speech discrimination in MS under "real world" conditions where sentences were presented in ecologically valid multi-talker speech or broadband noise at several signal-to-noise ratios (SNRs). MethodsPre-recorded Bamford-Kowal-Bench sentences were presented at five signal-to-noise ratios (SNR) in one of two background noises: speech-weighted noise and eight-talker babble. All auditory stimuli were presented via headphones to control (n = 38) and MS listeners with mild (n = 20), moderate (n = 16) and advanced (n = 10) disability. Disability was quantified by the Kurtzke Expanded Disability Status Scale (EDSS) and scored by a neurologist. All participants passed a routine audiometric examination. ResultsDespite normal hearing, MS psychometric discrimination curves which model the relationship between signal-to-noise ratio (SNR) and sentence discrimination accuracy in speech-weighted noise and babble did not change in slope (sentences/dB) but shifted to higher SNRs (dB) compared to controls. The magnitude of the shift in the curve systematically increased with greater disability. Furthermore, mixed-effects models identified EDSS score as the most significant predictor of speech discrimination in noise (odds ratio = 0.81; p < 0.001). Neither age, sex, disease phenotype or disease duration were significantly associated with speech discrimination performance in noise. Only MS listeners with advanced disability self-reported audio-attentional difficulty in a questionnaire designed to reflect auditory processing behaviours in daily life. ConclusionSpeech discrimination performance worsened systematically with greater disability, independent of age, sex, education, disease duration or disease phenotype. These results identify novel auditory processing deficits in MS and highlight that speech discrimination tasks may provide a viable non-invasive and sensitive means for disease monitoring in MS.

Noise-Sensitive But More Precise Subcortical Representations Coexist with Robust Cortical Encoding of Natural Vocalizations.

Humans and animals maintain accurate sound discrimination in the presence of loud sources of background noise. It is commonly assumed that this ability relies on the robustness of auditory cortex responses. However, only a few attempts have been made to characterize neural discrimination of communication sounds masked by noise at each stage of the auditory system and to quantify the noise effects on the neuronal discrimination in terms of alterations in amplitude modulations. Here, we measured neural discrimination between communication sounds masked by a vocalization-shaped stationary noise from multiunit responses recorded in the cochlear nucleus, inferior colliculus, auditory thalamus, and primary and secondary auditory cortex at several signal-to-noise ratios (SNRs) in anesthetized male or female guinea pigs. Masking noise decreased sound discrimination of neuronal populations in each auditory structure, but collicular and thalamic populations showed better performance than cortical populations at each SNR. In contrast, in each auditory structure, discrimination by neuronal populations was slightly decreased when tone-vocoded vocalizations were tested. These results shed new light on the specific contributions of subcortical structures to robust sound encoding, and suggest that the distortion of slow amplitude modulation cues conveyed by communication sounds is one of the factors constraining the neuronal discrimination in subcortical and cortical levels.SIGNIFICANCE STATEMENT Dissecting how auditory neurons discriminate communication sounds in noise is a major goal in auditory neuroscience. Robust sound coding in noise is often viewed as a specific property of cortical networks, although this remains to be demonstrated. Here, we tested the discrimination performance of neuronal populations at five levels of the auditory system in response to conspecific vocalizations masked by noise. In each acoustic condition, subcortical neurons better discriminated target vocalizations than cortical ones and in each structure, the reduction in discrimination performance was related to the reduction in slow amplitude modulation cues.

Interactions Between Digital Noise Reduction and Reverberation: Acoustic and Behavioral Effects.

Digital noise reduction (DNR) processing is used in hearing aids to enhance perception in noise by classifying and suppressing the noise acoustics. However, the efficacy of DNR processing is not known under reverberant conditions where the speech-in-noise acoustics are further degraded by reverberation. The purpose of this study was to investigate acoustic and perceptual effects of DNR processing across a range of reverberant conditions for individuals with hearing impairment. This study used an experimental design to investigate the effects of varying reverberation on speech-in-noise processed with DNR. Twenty-six listeners with mild-to-moderate sensorineural hearing impairment participated in the study. Speech stimuli were combined with unmodulated broadband noise at several signal-to-noise ratios (SNRs). A range of reverberant conditions with realistic parameters were simulated, as well as an anechoic control condition without reverberation. Reverberant speech-in-noise signals were processed using a spectral subtraction DNR simulation. Signals were acoustically analyzed using a phase inversion technique to quantify improvement in SNR as a result of DNR processing. Sentence intelligibility and subjective ratings of listening effort, speech naturalness, and background noise comfort were examined with and without DNR processing across the conditions. Improvement in SNR was greatest in the anechoic control condition and decreased as the ratio of direct to reverberant energy decreased. There was no significant effect of DNR processing on speech intelligibility in the anechoic control condition, but there was a significant decrease in speech intelligibility with DNR processing in all of the reverberant conditions. Subjectively, listeners reported greater listening effort and lower speech naturalness with DNR processing in some of the reverberant conditions. Listeners reported higher background noise comfort with DNR processing only in the anechoic control condition. Results suggest that reverberation affects DNR processing using a spectral subtraction algorithm in such a way that decreases the ability of DNR to reduce noise without distorting the speech acoustics. Overall, DNR processing may be most beneficial in environments with little reverberation and that the use of DNR processing in highly reverberant environments may actually produce adverse perceptual effects. Further research is warranted using commercial hearing aids in realistic reverberant environments.

Large-scale training to increase speech intelligibility for hearing-impaired listeners in novel noises.

Supervised speech segregation has been recently shown to improve human speech intelligibility in noise, when trained and tested on similar noises. However, a major challenge involves the ability to generalize to entirely novel noises. Such generalization would enable hearing aid and cochlear implant users to improve speech intelligibility in unknown noisy environments. This challenge is addressed in the current study through large-scale training. Specifically, a deep neural network (DNN) was trained on 10 000 noises to estimate the ideal ratio mask, and then employed to separate sentences from completely new noises (cafeteria and babble) at several signal-to-noise ratios (SNRs). Although the DNN was trained at the fixed SNR of - 2 dB, testing using hearing-impaired listeners demonstrated that speech intelligibility increased substantially following speech segregation using the novel noises and unmatched SNR conditions of 0 dB and 5 dB. Sentence intelligibility benefit was also observed for normal-hearing listeners in most noisy conditions. The results indicate that DNN-based supervised speech segregation with large-scale training is a very promising approach for generalization to new acoustic environments.

An algorithm to increase speech intelligibility for hearing-impaired listeners in entirely novel noises

Supervised speech segregation has been recently shown to improve human speech intelligibility in noise, when trained and tested on similar noises. However, a major challenge involves the ability to generalize to entirely novel noises. Such generalization would enable hearing aid andcochlear implant users to improve speech intelligibility in unknown noisy environments. This challenge is addressed in the current study through large-scale training. Specifically, a deep neural network (DNN) was trained on 10,000 noises to estimate the ideal ratio mask, and then employed to separate sentences from completely new noises (cafeteria and babble) at several signal-to-noise ratios (SNRs). Although the DNN wastrained at the fixed SNR of 2 dB, testing using hearing-impaired listeners demonstrated that speech intelligibility increased substantially following speech segregation using the novel noises and unmatched SNR conditions of 0 dB and 5 dB. Sentence-intelligibility benefit was also observed for normal-hearing listeners in most noisy conditions. The results indicate that DNN-based supervised speech segregation with large-scale training is a very promising approach for generalization to new acoustic environments.

Perception of conversational and clear speech syllables by native and non-native English-speaking listeners

In a recent study later, but not earlier, learners of English as a second language produced a smaller clear-speech benefit than native English-speaking talkers for vowels produced in six /bVd/ syllables (Rogers et al., 2010, JASA 123, 410–423). The present study compares perception of the same syllables by native and non-native English-speaking listeners. Conversational and clear-speech productions of the target syllables, “bead, bid, bayed, bed, bad,” and “bod,” were selected from three monolingual English speakers who had produced a significant clear-speech benefit in Rogers et al. (2010). The syllables were then mixed with noise at several signal-to-noise ratios (SNRs). Perception of these stimuli by three groups of listeners will be examined: (1) monolingual native English speakers, (2) ‘early’ learners of English as a second language, with an age of immersion (AOI) of 12 or earlier, and (3) later learners of English as a second language, with an AOI of 15 or later. Analyses of results of the six-alternative forced-choice task will focus on comparisons across listener groups, for the following measures: (1) estimates of clear-speech benefit at approximately 50% correct; (2) performance at a common SNR; and (3) estimates of the slope of the psychometric function. [Work supported by NIH.]

Improving word recognition in noise among hearing-impaired subjects with a single-channel cochlear noise-reduction algorithm

A common complaint of the hearing impaired is the inability to understand speech in noisy environments even with their hearing assistive devices. Only a few single-channel algorithms have significantly improved speech intelligibility in noise for hearing-impaired listeners. The current study introduces a cochlear noise reduction algorithm. It is based on a cochlear representation of acoustic signals and real-time derivation of a binary speech mask. The contribution of the algorithm for enhancing word recognition in noise was evaluated on a group of 42 normal-hearing subjects, 35 hearing-aid users, 8 cochlear implant recipients, and 14 participants with bimodal devices. Recognition scores of Hebrew monosyllabic words embedded in Gaussian noise at several signal-to-noise ratios (SNRs) were obtained with processed and unprocessed signals. The algorithm was not effective among the normal-hearing participants. However, it yielded a significant improvement in some of the hearing-impaired subjects under different listening conditions. Its most impressive benefit appeared among cochlear implant recipients. More than 20% improvement in recognition score of noisy words was obtained by 12, 16, and 26 hearing-impaired at SNR of 30, 24, and 18 dB, respectively. The algorithm has a potential to improve speech intelligibility in background noise, yet further research is required to improve its performances.

A nonlinear least-squares technique for analyzing probability of detection data that uses all the data without grouping

The purpose of a probability of detection (PD) experiment is to quantify the ability of a system to detect signals embedded in noise. The first step in the classical technique for analyzing data from such an experiment is to least-squares fit the data to a Gaussian probability integral using Gaussian probability paper or probits such that the Gaussian probability integral plots as a straight line. Because of the ease with which data can be fitted to a straight line, the classical technique does not require a modern digital computer. The classical technique has been effective over the years in many applications, however its shortcomings are (1) groups of data points must exist at several signal to noise ratios (SNRs) with about the same number of data points at each SNR and (2) groups with SNRs where detections occur 100 or 0% cannot be used. Data must be regrouped or thrown out to meet these criteria. The nonlinear least-squares (NLLS) technique described in this paper avoids these problems by least squares fitting the data to a Gaussian probability integral directly. This is now possible with modern digital computers. NLLS allows each data point to be used independently as a sample probability of detection of 1 or 0. This avoids the need for data to be grouped at all and permits using all the data. Results are compared with the classical technique, and favorable agreement is attained where all the conditions of the classical technique apply. Examples where all these conditions are not present are also provided, and the advantage of the new technique is demonstrated. The analysis is performed with a program written in basic and run on a home computer. In applying the new technique to the examples, it was noted that errors in quantifying the ability of a system to detect signals can be large. Further analysis showed that the size of the errors depended on the location of the SNRs, and an iterative method is demonstrated for selecting SNRs such that reasonable accuracy is obtained with a small number (50) of data samples.

Alignment methods for averaging of high-resolution cardiac signals: a comparative study of performance

Accurate signal estimation by means of coherent averaging techniques needs temporal alignment methods. A known low-pass filtering effect is yielded when alignment errors are present. This is very critical in the estimation of low-level high-frequency potentials in high-resolution ECG analysis. A comparative study of the performance of three alignment methods (the double-level method, a new time-delay estimation method based on normalized integrals, and matched filtering) is presented in this paper. A real signal and additive random noise for several signal-to-noise ratios (SNR's) are selected to make an ensemble of computer-simulated beats. The relation between the standard deviation of temporal misalignment versus SNR is discussed. A second study with real ECG signals is also presented. Several morphologies of QRS and P waves are tested. The results are in agreement with the computer simulation study. Nevertheless, the power spectrum of the noise process can affect the results. Matched filter estimation has been tested in the presence of power line interference (50 Hz), with poor results. An application of the three alignment methods as a function of the SNR is proposed. The new time-delay estimation method has been observed to be robust, even in the presence of nonwhite noise.