Auditory filter-bank compression improves estimation of signal-to-noise ratio for speech in noise.

Signal-to-noise ratio (SNR) estimation is necessary for many speech processing applications often challenged by nonstationary noise. The authors have previously demonstrated that the variance of spectral entropy (VSE) is a reliable estimate of SNR in nonstationary noise. Based on pre-estimated VSE-SNR relationship functions, the SNR of unseen acoustic environments can be estimated from the measured VSE. This study predicts that introducing a compressive function based on cochlear processing will increase the stability of the pre-estimated VSE-SNR relationship functions. This study demonstrates that calculating the VSE based on a nonlinear filter-bank, simulating cochlear compression, reduces the VSE-based SNR estimation errors. VSE-SNR relationship functions were estimated using speech tokens presented in babble noise comprised of different numbers of speakers. Results showed that the coefficient of determination (R2) of the estimated VSE-SNR relationship functions have absolute percentage improvements of over 26% when using a filter-bank with a compressive function, compared to when using a linear filter-bank without compression. In 2-talker babble noise, the estimation accuracy is more than 3 dB better than other published methods.

Effect of auditory efferent time-constant duration on speech recognition in noise.

The human auditory efferent system may play a role in improving speech-in-noise recognition with an associated range of time constants. Computational auditory models with efferent-inspired feedback demonstrate improved speech-in-noise recognition with long efferent time constants (2000 ms). This study used a similar model plus an Automatic Speech Recognition (ASR) system to investigate the role of shorter time constants. ASR speech recognition in noise improved with efferent feedback (compared to no-efferent feedback) for both short and long efferent time constants. For some signal-to-noise ratios, speech recognition in noise improved as efferent time constants were increased from 118 to 2000 ms.

Effect of duration and gating of the signal on the binaural masking level difference for narrowband and broadband maskers.

Thresholds were measured for a 250-Hz signal with an interaural phase difference of 0° (diotic) or 180° (dichotic), with signal durations of 12 and 60 ms (including 6-ms ramps) and 300 ms (including 6- or 50-ms ramps). The signal-centered diotic noise masker had a bandwidth of 20 or 200 Hz. For the 20-Hz wide masker, the binaural masking level difference (BMLD), i.e., threshold difference between diotic and dichotic signal, increased with signal duration and, for the 300-ms signal, the BMLD was larger with 50-ms rather than 6-ms ramps. These signal parameters hardly affected the BMLD for the 200-Hz wide masker.

Modeling off-frequency binaural masking for short- and long-duration signals.

Experimental binaural masking-pattern data are presented together with model simulations for 12- and 600-ms signals. The masker was a diotic 11-Hz wide noise centered on 500 Hz. The tonal signal was presented either diotically or dichotically (180° interaural phase difference) with frequencies ranging from 400 to 600 Hz. The results and the modeling agree with previous data and hypotheses; simulations with a binaural model sensitive to monaural modulation cues show that the effect of duration on off-frequency binaural masking-level differences is mainly a result of modulation cues which are only available in the monaural detection of long signals.

Frequency Tuning of the Efferent Effect on Cochlear Gain in Humans.

Cochlear gain reduction via efferent feedback from the medial olivocochlear bundle is frequency specific (Guinan, Curr Opin Otolaryngol Head Neck Surg 18:447-453, 2010). The present study with humans used the Fixed Duration Masking Curve psychoacoustical method (Yasin et al., J Acoust Soc Am 133:4145-4155, 2013a; Yasin et al., Basic aspects of hearing: physiology and perception, pp 39-46, 2013b; Yasin et al., J Neurosci 34:15319-15326, 2014) to estimate the frequency specificity of the efferent effect at the cochlear level. The combined duration of the masker-plus-signal stimulus was 25 ms, within the efferent onset delay of about 31-43 ms (James et al., Clin Otolaryngol 27:106-112, 2002). Masker level (4.0 or 1.8 kHz) at threshold was obtained for a 4-kHz signal in the absence or presence of an ipsilateral 60 dB SPL, 160-ms precursor (200-Hz bandwidth) centred at frequencies between 2.5 and 5.5 kHz. Efferent-mediated cochlear gain reduction was greatest for precursors with frequencies the same as, or close to that of, the signal (gain was reduced by about 20 dB), and least for precursors with frequencies well removed from that of the signal (gain remained at around 40 dB). The tuning of the efferent effect filter (tuning extending 0.5-0.7 octaves above and below the signal frequency) is within the range obtained in humans using otoacoustic emissions (Lilaonitkul and Guinan, J Assoc Res Otolaryngol 10:459-470, 2009; Zhao and Dhar, J Neurophysiol 108:25-30, 2012). The 10 dB bandwidth of the efferent-effect filter at 4000 Hz was about 1300 Hz (Q(10) of 3.1). The FDMC method can be used to provide an unbiased measure of the bandwidth of the efferent effect filter using ipsilateral efferent stimulation.

Neural correlates of attention and streaming in a perceptually multistable auditory illusion.

In a complex acoustic environment, acoustic cues and attention interact in the formation of streams within the auditory scene. In this study, a variant of the "octave illusion" [Deutsch (1974). Nature 251, 307-309] was used to investigate the neural correlates of auditory streaming, and to elucidate the effects of attention on the interaction between sequential and concurrent sound segregation in humans. By directing subjects' attention to different frequencies and ears, it was possible to elicit several different illusory percepts with the identical stimulus. The first experiment tested the hypothesis that the illusion depends on the ability of listeners to perceptually stream the target tones from within the alternating sound sequences. In the second experiment, concurrent psychophysical measures and electroencephalography recordings provided neural correlates of the various percepts elicited by the multistable stimulus. The results show that the perception and neural correlates of the auditory illusion can be manipulated robustly by attentional focus and that the illusion is constrained in much the same way as auditory stream segregation, suggesting common underlying mechanisms.

Effect of human auditory efferent feedback on cochlear gain and compression.

The mammalian auditory system includes a brainstem-mediated efferent pathway from the superior olivary complex by way of the medial olivocochlear system, which reduces the cochlear response to sound (Warr and Guinan, 1979; Liberman et al., 1996). The human medial olivocochlear response has an onset delay of between 25 and 40 ms and rise and decay constants in the region of 280 and 160 ms, respectively (Backus and Guinan, 2006). Physiological studies with nonhuman mammals indicate that onset and decay characteristics of efferent activation are dependent on the temporal and level characteristics of the auditory stimulus (Bacon and Smith, 1991; Guinan and Stankovic, 1996). This study uses a novel psychoacoustical masking technique using a precursor sound to obtain a measure of the efferent effect in humans. This technique avoids confounds currently associated with other psychoacoustical measures. Both temporal and level dependency of the efferent effect was measured, providing a comprehensive measure of the effect of human auditory efferents on cochlear gain and compression. Results indicate that a precursor (>20 dB SPL) induced efferent activation, resulting in a decrease in both maximum gain and maximum compression, with linearization of the compressive function for input sound levels between 50 and 70 dB SPL. Estimated gain decreased as precursor level increased, and increased as the silent interval between the precursor and combined masker-signal stimulus increased, consistent with a decay of the efferent effect. Human auditory efferent activation linearizes the cochlear response for mid-level sounds while reducing maximum gain.

Improved psychophysical methods to estimate peripheral gain and compression.

It is possible that previous psychophysical estimates of basilar membrane gain and compression using temporal masking curve (TMC) and additivity of forward masking (AFM) methods using long-duration maskers (>30 ms) could have been affected by activation of the medial olivocochlear reflex (MOCR) (Jennings et al. ; Plack and Arifianto ). In experiment 1, AFM and TMC methods were compared to a new fixed-duration masking curve (FDMC) method in which the combined masker and signal stimulus duration is fixed at 25 ms. Estimates of compression were found to be not significantly different for TMC, FDMC and AFM methods. Estimates of gain were similar for TMC and FDMC methods. Maximum compression was associated with a significantly lower input masker level using the FDMC compared to the TMC method. In experiment 2, the FDMC method was used to investigate the effect of efferent activation on gain and compression estimates by presenting a precursor sound prior to the combined masker-signal stimulus. Estimated gain decreased as precursor level increased, and increased as the silent interval between the precursor and combined masker-signal stimulus increased, consistent with a decay of the efferent response.

Estimating peripheral gain and compression using fixed-duration masking curves.

Estimates of human basilar membrane gain and compression obtained using temporal masking curve (TMC) and additivity of forward masking (AFM) methods with long-duration maskers or long masker-signal silent intervals may be affected by olivocochlear efferent activation, which reduces basilar membrane gain. The present study introduces a fixed-duration masking curve (FDMC) method, which involves a comparison of off- and on-frequency forward masker levels at threshold as a function of masker and signal duration, with the total masker-signal duration fixed at 25 ms to minimize efferent effects. Gain and compression estimates from the FDMC technique were compared with those from TMC (104-ms maskers) and AFM (10- and 200-ms maskers) methods. Compression estimates over an input-masker range of 40-60 dB sound pressure level were similar for the four methods. Maximum compression occurred at a lower input level for the FDMC compared to the TMC method. Estimates of gain were similar for TMC and FDMC methods. The FDMC method may provide a more reliable estimate of BM gain and compression in the absence of efferent activation and could be a useful method for estimating effects of efferent activity when used with a precursor sound (to trigger efferent activation), presented prior to the combined masker-signal stimulus.

The effects of noise-bandwidth, noise-fringe duration, and temporal signal location on the binaural masking-level difference.

The effects of forward and backward noise fringes on binaural signal detectability were investigated. Masked thresholds for a 12-ms, 250-Hz, sinusoidal signal masked by Gaussian noise, centered at 250 Hz, with bandwidths from 3 to 201 Hz, were obtained in N(0)S(0) and N(0)S(π) configurations. The signal was (a) temporally centered in a 12-ms noise burst (no fringe), (b) presented at the start of a 600-ms noise burst (backward fringe), or (c) temporally centered in a 600-ms noise burst (forward-plus-backward fringe). For noise bandwidths between 3 and 75 Hz, detection in N(0)S(0) improved with the addition of a backward fringe, improving further with an additional forward fringe; there was little improvement in N(0)S(π). The binaural masking-level difference (BMLD) increased from 0 to 8 dB with a forward-plus-backward fringe as noise bandwidths increased to 100 Hz, increasing slightly to 10 dB at 201 Hz. This two-stage increase was less pronounced with a backward fringe. With no fringe, the BMLD was about 10-14 dB at all bandwidths. Performance appears to result from the interaction of across-time and across-frequency listening strategies and the possible effects of gain reduction and suppression, which combine in complex ways. Current binaural models are, as yet, unable to account fully for these effects.

Correspondence between three-dimensional ear depth information derived from two-dimensional images and magnetic resonance imaging: Use of a neural-network model.

There is much interest in anthropometric-derived head-related transfer functions (HRTFs) for simulating audio for virtual-reality systems. Three-dimensional (3D) anthropometric measures can be measured directly from individuals, or indirectly simulated from two-dimensional (2D) pinna images. The latter often requires additional pinna, head and/or torso measures. This study investigated accuracy with which 3D depth information can be obtained solely from 2D pinna images using an unsupervised monocular-depth estimation neural-network model. Output was compared to depth information obtained from corresponding magnetic resonance imaging (MRI) head scans (ground truth). Results show that 3D depth estimates obtained from 2D pinna images corresponded closely with MRI head-scan depth values.

Speech versus tone processing in compensated dyslexia: discrimination and lateralization with a dichotic mismatch negativity (MMN) paradigm.

This study investigated the possibility of a pervasive auditory-processing deficit in 10 adult dyslexics who had compensated for their reading disability, compared to 10 matched controls. Unlike previous studies [Baldeweg, T., Richardson, A., Watkins, S., Foale, C. & Gruzelier, J. (1999). Impaired auditory frequency discrimination in dyslexia detected with mismatch evoked potentials. Annals of Neurology, 45(4): 495-503], the current EEG study used a dichotic presentation of stimuli in order to probe the relationship between ear advantage and left- and right-hemisphere dominance for processing speech and non-speech stimuli respectively. A dichotic presentation is thought to maximise lateralization effects, as well being a more ecologically valid paradigm. The Mismatch Negativity (MMN) was measured for both speech [consonant-vowel pairs: /ta/ vs. /ka/ and ba/ vs. /da/] and non-speech stimuli (pure tones: 1 kHz vs. 1.2 kHz). Smaller MMNs to tone stimuli were obtained for dyslexics versus controls, but no differences were found for speech stimuli. Controls differentiated between speech and tones, with larger MMNs to tone stimuli. Dyslexics showed significantly greater MMNs to one stop consonant discrimination (/ta/ vs /ka/) than the other (/ba/ vs /da/), but did not differentiate speech from tones, and no or minimal lateralization was found for either group or stimulus type, in line with recent studies [Kershner and Micallef, 1992; Bellis, T.J., Nicol, T., & Kraus, N. (2000). Aging affects hemispheric asymmetry in the neural representation of speech sounds. Journal of Neuroscience, 20, 791-797]. However, analysis of left and right dipole source activity suggested right hemispheric preference for tones in controls, and the lack of such a preference in dyslexics. Lateralization of the auditory system in general may be less specialized in compensated dyslexia, although no specific differences in speech lateralization were seen. The present study also extends previous findings to show that the frequency range over which dyslexics are shown to display impaired frequency discrimination can be extended to 20% (1 kHz vs. 1.2 kHz) if using a dichotic presentation.

Duration of auditory sensory memory in parents of children with SLI: a mismatch negativity study.

In a previous behavioral study, we showed that parents of children with SLI had a subclinical deficit in phonological short-term memory. Here, we tested the hypothesis that they also have a deficit in nonverbal auditory sensory memory. We measured auditory sensory memory using a paradigm involving an electrophysiological component called the mismatch negativity (MMN). The MMN is a measure of the brain's ability to detect a difference between a frequent standard stimulus (1000 Hz tone) and a rare deviant one (1200 Hz tone). Memory effects were assessed by varying the inter-stimulus interval (ISI) between the standard and deviant. We predicted that parents of children with SLI would have a smaller MMN than parents of typically developing children at a long ISI (3000 ms), but not at a short one (800 ms). This was broadly confirmed. However, individual differences in MMN amplitude did not correlate with measures of phonological short-term memory. Attenuation of MMN amplitude at the longer ISI thus did not provide unambiguous support for the hypothesis of a reduced auditory sensory memory in parents of affected children. We conclude by reviewing possible explanations for the observed group effects.

Hemispheric differences in processing dichotic meaningful and non-meaningful words.

Classic dichotic-listening paradigms reveal a right-ear advantage (REA) for speech sounds as compared to non-speech sounds. This REA is assumed to be associated with a left-hemisphere dominance for meaningful speech processing. This study objectively probed the relationship between ear advantage and hemispheric dominance in a dichotic-listening situation, using event-related potentials (ERPs). The mismatch negativity (MMN) and a late negativity (LN) were measured for bisyllabic meaningful words and non-meaningful pseudowords, which differed in their second syllable. Eighteen normal-hearing listeners were presented with a repeating diotic standard ([beI-gi:] or [leI-gi:]) and an occasional dichotic deviant (a standard presented to one ear and a deviant [beI-bi:], [beI-di:], [leI-bi:] or [leI-di:] presented to the opposite ear). As predicted there was a REA for meaningful words compared to non-meaningful words. Also, dipole source analysis suggested that dipole strength was stronger in the left than the right cortical region for meaningful words. However, there were differences in response within meaningful words as well as between meaningful and non-meaningful words which may be explained by the characteristics of embedded words and the position-specific probability of phoneme occurrence in words.

Effect of efferent activation on binaural frequency selectivity.

Binaural notched-noise experiments indicate a reduced frequency selectivity of the binaural system compared to monaural processing. The present study investigates how auditory efferent activation (via the medial olivocochlear system) affects binaural frequency selectivity in normal-hearing listeners. Thresholds were measured for a 1-kHz signal embedded in a diotic notched-noise masker for various notch widths. The signal was either presented in phase (diotic) or in antiphase (dichotic), gated with the noise. Stimulus duration was 25 ms, in order to avoid efferent activation due to the masker or the signal. A bandpass-filtered noise precursor was presented prior to the masker and signal stimuli to activate the efferent system. The silent interval between the precursor and the masker-signal complex was 50 ms. For comparison, thresholds for detectability of the masked signal were also measured in a baseline condition without the precursor and, in addition, without the masker. On average, the results of the baseline condition indicate an effectively wider binaural filter, as expected. For both signal phases, the addition of the precursor results in effectively wider filters, which is in agreement with the hypothesis that cochlear gain is reduced due to the presence of the precursor.

An auditory illusion reveals the role of streaming in the temporal misallocation of perceptual objects.

This study investigates the neural correlates and processes underlying the ambiguous percept produced by a stimulus similar to Deutsch's 'octave illusion', in which each ear is presented with a sequence of alternating pure tones of low and high frequencies. The same sequence is presented to each ear, but in opposite phase, such that the left and right ears receive a high-low-high … and a low-high-low … pattern, respectively. Listeners generally report hearing the illusion of an alternating pattern of low and high tones, with all the low tones lateralized to one side and all the high tones lateralized to the other side. The current explanation of the illusion is that it reflects an illusory feature conjunction of pitch and perceived location. Using psychophysics and electroencephalogram measures, we test this and an alternative hypothesis involving synchronous and sequential stream segregation, and investigate potential neural correlates of the illusion. We find that the illusion of alternating tones arises from the synchronous tone pairs across ears rather than sequential tones in one ear, suggesting that the illusion involves a misattribution of time across perceptual streams, rather than a misattribution of location within a stream. The results provide new insights into the mechanisms of binaural streaming and synchronous sound segregation.This article is part of the themed issue 'Auditory and visual scene analysis'.

The role of suppression in the upward spread of masking.

The upward spread of masking refers to the higher growth rate of masking for maskers lower in frequency than the signal, compared to maskers at the signal frequency (Wegel RL, Lane CE. The auditory masking of one pure tone by another and its possible relation to the dynamics of the inner ear. Physics Rev. 23:266-285, 1924; Egan JP, Hake HW. On the masking pattern of a simple auditory stimulus. J. Acoust. Soc. Am. 22:622-630, 1950; Delgutte B. Physiological mechanisms of psychophysical masking: Observations from auditory-nerve fibres. J. Acoust. Soc. Am. 87:791-809, 1990a, Delgutte B. Two-tone rate suppression in auditory-nerve fibres: Dependence on suppressor frequency and level. Hear Res. 49:225-246, 1990b). The upward spread of simultaneous masking may arise from a combination of excitatory and suppressive effects. In this study, growth of masking functions were obtained for a 4-kHz signal masked by an on-frequency (4 kHz) or off-frequency (2.4 kHz), simultaneous or forward masker, in the presence of a notched noise with a center frequency of 4 kHz presented to restrict off-frequency listening. Compression was estimated from the slopes of the off-frequency growth of masking functions. Suppression was estimated by comparing the off-frequency simultaneous- and forward-masked growth of masking functions. Results showed that, for midlevel signals (35-60 dB SPL), the compression exponent estimated from simultaneous and forward masking averaged 0.31 and 0.26, respectively. The maximum amount of suppression (defined as the decrease in the basilar-membrane response to the signal) was variable, ranging from about 6 to 17 dB across subjects. Despite the substantial reduction in the response to the signal, the results suggest that suppression has a minimal effect on the slope of the masking function at mid levels. Rather, upward spread of masking seems to be mainly determined by the compressive basilar-membrane response to the signal in relation to the linear response to the lower-frequency masker.

Objective measures of binaural masking level differences and comodulation masking release based on late auditory evoked potentials.

The audibility of important sounds is often hampered due to the presence of other masking sounds. The present study investigates if a correlate of the audibility of a tone masked by noise is found in late auditory evoked potentials measured from human listeners. The audibility of the target sound at a fixed physical intensity is varied by introducing auditory cues of (i) interaural target signal phase disparity and (ii) coherent masker level fluctuations in different frequency regions. In agreement with previous studies, psychoacoustical experiments showed that both stimulus manipulations result in a masking release (i: binaural masking level difference; ii: comodulation masking release) compared to a condition where those cues are not present. Late auditory evoked potentials (N1, P2) were recorded for the stimuli at a constant masker level, but different signal levels within the same set of listeners who participated in the psychoacoustical experiment. The data indicate differences in N1 and P2 between stimuli with and without interaural phase disparities. However, differences for stimuli with and without coherent masker modulation were only found for P2, i.e., only P2 is sensitive to the increase in audibility, irrespective of the cue that caused the masking release. The amplitude of P2 is consistent with the psychoacoustical finding of an addition of the masking releases when both cues are present. Even though it cannot be concluded where along the auditory pathway the audibility is represented, the P2 component of auditory evoked potentials is a candidate for an objective measure of audibility in the human auditory system.

Effects of sequential streaming on auditory masking using psychoacoustics and auditory evoked potentials.

The present study was aimed at investigating the relationship between the mismatch negativity (MMN) and psychoacoustical effects of sequential streaming on comodulation masking release (CMR). The influence of sequential streaming on CMR was investigated using a psychoacoustical alternative forced-choice procedure and electroencephalography (EEG) for the same group of subjects. The psychoacoustical data showed, that adding precursors comprising of only off-signal-frequency maskers abolished the CMR. Complementary EEG data showed an MMN irrespective of the masker envelope correlation across frequency when only the off-signal-frequency masker components were present. The addition of such precursors promotes a separation of the on- and off-frequency masker components into distinct auditory objects preventing the auditory system from using comodulation as an additional cue. A frequency-specific adaptation changing the representation of the flanking bands in the streaming conditions may also contribute to the reduction of CMR in the stream conditions, however, it is unlikely that adaptation is the primary reason for the streaming effect. A neurophysiological correlate of sequential streaming was found in EEG data using MMN, but the magnitude of the MMN was not correlated with the audibility of the signal in CMR experiments. Dipole source analysis indicated different cortical regions involved in processing auditory streaming and modulation detection. In particular, neural sources for processing auditory streaming include cortical regions involved in decision-making.

The effects of low- and high-frequency suppressors on psychophysical estimates of basilar-membrane compression and gain.

Physiological studies suggest that the increase in suppression as a function of suppressor level is greater for a suppressor below than above the signal frequency. This study investigated the pattern of gain reduction underlying this increase in suppression. Temporal masking curves (TMCs) were obtained by measuring the level of a 2.2-kHz sinusoidal off-frequency masker or 4-kHz on-frequency sinusoidal masker required to mask a brief 4-kHz sinusoidal signal at 10 dB SL, for masker-signal intervals of 20-100 ms. TMCs were also obtained in the presence of a 3- or 4.75-kHz sinusoidal suppressor gated with the 4-kHz masker, for suppressor levels of 40-70 dB SPL. The decrease in gain (increase in suppression) as a function of suppressor level was greater with a 3-kHz suppressor than with a 4.75-kHz suppressor, in line with previous findings. Basilar membrane input-output (I/O) functions derived from the TMCs showed a shift to higher input (4-kHz masker) levels of the low-level (linear) portion of the I/O function with the addition of a suppressor, with partial linearization of the function, but no reduction in maximum compression.