Intelligibility Of Sentences In Noise Research Articles

The effect of phoneme-based auditory training on speech intelligibility in hearing-aid users

Objective Hearing loss commonly causes difficulties in understanding speech in the presence of background noise. The benefits of hearing-aids in terms of speech intelligibility in challenging listening scenarios remain limited. The present study investigated if phoneme-in-noise discrimination training improves phoneme identification and sentence intelligibility in noise in hearing-aid users. Design Two groups of participants received either a two-week training program or a control intervention. Three phoneme categories were trained: onset consonants (C1), vowels (V) and post-vowel consonants (C2) in C1-V-C2-/i/ logatomes from the Danish nonsense word corpus (DANOK). Phoneme identification test and hearing in noise test (HINT) were administered before and after the respective interventions and, for the training group only, after three months. Study sample Twenty 63-to-79 years old individuals with a mild-to-moderate sensorineural hearing loss and at least one year of experience using hearing-aids. Results The training provided an improvement in phoneme identification scores for vowels and post-vowel consonants, which was retained over three months. No significant performance improvement in HINT was found. Conclusion The study demonstrates that the training induced a robust refinement of auditory perception at a phoneme level but provides no evidence for the generalisation to an untrained sentence intelligibility task.

Hearing in Real-Life Environments (HERE): Structure and Reliability of a Questionnaire on Perceived Hearing for Older Adults.

The ability to hear in a variety of social situations and environments is vital for social participation and a high quality of life. One way to assess hearing ability is by means of self-report questionnaire. For questionnaires to be useful, their measurement properties, based on careful validation, have to be known. Only recently has consensus been reached concerning how to perform such validation and been published as COSMIN (consensus-based standards for the selection of health status measurement instruments) guidelines. Here the authors use these guidelines to evaluate the measurement properties of the "Hearing in Real-Life Environments" (HERE) questionnaire, a newly developed self-report measure that assesses speech perception, spatial orientation, and the social-emotional consequences of hearing impairment in older adults. The aim is to illustrate the process of validation and encourage similar examinations of other frequently used questionnaires. The HERE questionnaire includes 15 items with a numeric rating scale from 0 to 10 for each item and allows the assessment of hearing with and without hearing aids. The evaluation was performed in two cohorts of community-dwelling older adults from Finland (n = 581, mean 82 years) and the United Kingdom (n = 50, mean 69 years). The internal structure of the questionnaire and its relationship to age, hearing level, and self-reported and behavioral measures of speech perception was assessed and, when possible, compared between cohorts. The results of the factor analysis showed that the HERE's internal structure was similar across cohorts. In both cohorts, the factor analysis showed a satisfactory solution for three factors (speech hearing, spatial hearing, and socio-emotional consequences), with a high internal consistency for each factor (Cronbach's α's for the factors from 0.90 to 0.97). Test-retest analysis showed the HERE overall mean score to be stable and highly replicable over time (intraclass correlation coefficient = 0.86, standard error of measurement of the test score = 0.92). The HERE overall mean score correlated highly with another self-report measure of speech perception, the Speech Spatial Qualities of Hearing questionnaire (standardized regression coefficient [β] = -0.75, p < 0.001), moderately highly with behaviorally assessed hearing level (best-ear average: β = 0.45 to 0.46), and moderately highly with behaviorally measured intelligibility of sentences in noise (β = -0.50, p < 0.001). Using the COSMIN guidelines, the authors show that the HERE is a valid, reliable, and stable questionnaire for the assessment of self-reported speech perception, sound localization, and the socio-emotional consequences of hearing impairment in the context of social functioning. The authors also show that cross-cultural data collected using different data collection strategies can be combined with a range of statistical methods to validate a questionnaire.

Speech intelligibility of virtual humans

The speech intelligibility benefit of visual speech cues during oral communication is well-established. Therefore, an ecologically valid approach of auditory assessment should include the processing of both auditory and visual speech cues. This study describes the development and evaluation of a virtual human speaker designed to present speech auditory-visually. A male and female virtual human speaker were created and evaluated in two experiments: a visual-only speech reading test of words and sentences and an auditory-visual speech intelligibility sentence test. A group of five hearing, skilled speech reading adults participated in the speech reading test whereas a group of young normal-hearing participants (N = 35) was recruited for the intelligibility test. Skilled speech readers correctly identified 57 to 67% of the words and sentences uttered by the virtual speakers. The presence of the virtual speaker improved the speech intelligibility of sentences in noise by 1.5 to 2 dB. These results demonstrate the potential applicability of virtual humans in future auditory-visual speech assessment paradigms.

Evaluation of a Frequency-Lowering Algorithm for Adults With High-Frequency Hearing Loss.

The objective was to determine the effects of a frequency-lowering algorithm (frequency composition, Fcomp) on consonant identification, word-final /s, z/ detection, the intelligibility of sentences in noise, and subjective benefit, for people with high-frequency hearing loss, including people with dead regions (DRs) in the cochlea. A single-blind randomized crossover design was used. Performance with Bernafon Acriva 9 hearing aids was compared with Fcomp off and Fcomp on. Participants wore the hearing aids in each condition in a counterbalanced order. Data were collected after at least 8 weeks of experience with a condition. Outcome measures were audibility, scores from the speech perception tests, and scores from a questionnaire comparing self-perceived hearing ability with Fcomp off and Fcomp on. Ten adults with mild to severe high-frequency hearing loss (seven with extensive DRs, one with patchy or restricted DRs, and two with no DR) were tested. Fcomp improved the audibility of high-frequency sounds for 6 out of 10 participants. There was no overall effect of Fcomp on consonant identification, but the pattern of consonant confusions varied across conditions and participants. For word-final /s, z/ detection, performance was significantly better with Fcomp on than with Fcomp off. Questionnaire scores showed no differences between conditions. In summary, Fcomp improved word-final /s, z/ detection. No benefit was found for the other measures.

An algorithm to improve speech recognition in noise for hearing-impaired listeners: Consonant identification and articulatory feature transmission

Previous work has shown that a supervised-learning algorithm estimating the ideal binary mask (IBM) can improve sentence intelligibility in noise for hearing-impaired (HI) listeners from scores below 30% to above 80% [Healy et al., J. Acoust. Soc. Am. 134 (2013)]. The algorithm generates a binary mask by using a deep neural network to classify speech-dominant and noise-dominant time-frequency units. In the current study, these results are extended to consonant recognition, in order to examine the specific speech cues responsible for the observed performance improvements. Consonant recognition in speech-shaped noise or babble was examined in normal-hearing and HI listeners in three conditions: unprocessed, noise removed via the IBM, and noise removed via the classification-based algorithm. The IBM demonstrated substantial performance improvements, averaging up to 45% points. The algorithm also produced sizeable gains, averaging up to 34% points. An information-transmission analysis of cues associated with manner of articulation, place of articulation, and voicing indicated general similarity in the cues transmitted by the IBM versus the algorithm. However, important differences were observed, which may guide the future refinement of the algorithm. [Work supported by NIH.]

Predicting the speech reception thresholds with physical metrics

Many measures [e.g., speech transmission index (STI) and speech intelligibility index (SII)] have been proposed to predict the speech intelligibility in noise. Nevertheless, most of these studies were performed under the conditions with a limited number of maskers. The present study further investigated how well the present speech intelligibility and quality metrics predicted the speech reception thresholds (SRTs) for sentences corrupted by stationary and fluctuating maskers. The SRT scores were collected from 30 normal-hearing (NH) and 15 hearing-impaired (HI) native-Cantonese listeners. Sentences were corrupted by nine types of maskers, including speech-shaped noise and eight real-life environmental noises (4- and 6-talker babbles, upper and lower deck in bus, cafe, Chinese restaurant, MTR carriage, and street). The resulting average SRT scores were subject to the correlation analysis with various metrics computed from the noise-masked sentences. Of all the objective metrics considered, the STI and CSII measures performed the best, and their high correlations (i.e., r = 0.91 to 0.96) were maintained in both NH and HI conditions. This suggests that some of the physical metrics that have been found previously to correlate highly with the intelligibility of sentences in noise may also be used to predict the SRTs affected by different maskers.

Spectrotemporal Modulation Sensitivity as a Predictor of Speech Intelligibility for Hearing-Impaired Listeners

A model that can accurately predict speech intelligibility for a given hearing-impaired (HI) listener would be an important tool for hearing-aid fitting or hearing-aid algorithm development. Existing speech-intelligibility models do not incorporate variability in suprathreshold deficits that are not well predicted by classical audiometric measures. One possible approach to the incorporation of such deficits is to base intelligibility predictions on sensitivity to simultaneously spectrally and temporally modulated signals. The likelihood of success of this approach was evaluated by comparing estimates of spectrotemporal modulation (STM) sensitivity to speech intelligibility and to psychoacoustic estimates of frequency selectivity and temporal fine-structure (TFS) sensitivity across a group of HI listeners. The minimum modulation depth required to detect STM applied to an 86 dB SPL four-octave noise carrier was measured for combinations of temporal modulation rate (4, 12, or 32 Hz) and spectral modulation density (0.5, 1, 2, or 4 cycles/octave). STM sensitivity estimates for individual HI listeners were compared to estimates of frequency selectivity (measured using the notched-noise method at 500, 1000, 2000, and 4000 Hz), TFS processing ability (2 Hz frequency-modulation detection thresholds for 500, 1000, 2000, and 4000 Hz carriers) and sentence intelligibility in noise (at a 0 dB signal-to-noise ratio) that were measured for the same listeners in a separate study. Eight normal-hearing (NH) listeners and 12 listeners with a diagnosis of bilateral sensorineural hearing loss participated. STM sensitivity was compared between NH and HI listener groups using a repeated-measures analysis of variance. A stepwise regression analysis compared STM sensitivity for individual HI listeners to audiometric thresholds, age, and measures of frequency selectivity and TFS processing ability. A second stepwise regression analysis compared speech intelligibility to STM sensitivity and the audiogram-based Speech Intelligibility Index. STM detection thresholds were elevated for the HI listeners, but only for low rates and high densities. STM sensitivity for individual HI listeners was well predicted by a combination of estimates of frequency selectivity at 4000 Hz and TFS sensitivity at 500 Hz but was unrelated to audiometric thresholds. STM sensitivity accounted for an additional 40% of the variance in speech intelligibility beyond the 40% accounted for by the audibility-based Speech Intelligibility Index. Impaired STM sensitivity likely results from a combination of a reduced ability to resolve spectral peaks and a reduced ability to use TFS information to follow spectral-peak movements. Combining STM sensitivity estimates with audiometric threshold measures for individual HI listeners provided a more accurate prediction of speech intelligibility than audiometric measures alone. These results suggest a significant likelihood of success for an STM-based model of speech intelligibility for HI listeners.

The Influence of Semantically Related and Unrelated Text Cues on the Intelligibility of Sentences in Noise

In two experiments with different subject groups, we explored the relationship between semantic context and intelligibility by examining the influence of visually presented, semantically related, and unrelated three-word text cues on perception of spoken sentences in stationary noise across a range of speech-to-noise ratios (SNRs). In addition, in Experiment (Exp) 2, we explored the relationship between individual differences in cognitive factors and the effect of the cues on speech intelligibility. In Exp 1, cues had been generated by participants themselves in a previous test session (own) or by someone else (alien). These cues were either appropriate for that sentence (match) or for a different sentence (mismatch). A condition with nonword cues, generated by the experimenter, served as a control. Experimental sentences were presented at three SNRs (dB SNR) corresponding to the entirely correct repetition of 29%, 50%, or 71% of sentences (speech reception thresholds; SRTs). In Exp 2, semantically matching or mismatching cues and nonword cues were presented before sentences at SNRs corresponding to SRTs of 16% and 29%. The participants in Exp 2 also performed tests of verbal working memory capacity and the ability to read partially masked text. In Exp 1, matching cues improved perception relative to the nonword and mismatching cues, with largest benefits at the SNR corresponding to 29% performance in the SRT task. Mismatching cues did not impair speech perception relative to the nonword cue condition, and no difference in the effect of own and alien matching cues was observed. In Exp 2, matching cues improved speech perception as measured using both the percentage of correctly reported words and the percentage of entirely correctly reported sentences. Mismatching cues reduced the percentage of repeated words (but not the sentence-based scores) compared with the nonword cue condition. Working memory capacity and ability to read partly masked sentences were positively associated with the number of sentences repeated entirely correctly in the mismatch condition at the 29% SNR. In difficult listening conditions, both relevant and irrelevant semantic context can influence speech perception in noise. High working memory capacity and good linguistic skills are associated with a greater ability to inhibit irrelevant context when uncued sentence intelligibility is around 29% correct.

Results with a cochlear implant channel-picking strategy based on “Selected Groups”

A novel channel-picking strategy for cochlear implants (CIs) which considers the spatial distribution and the spectral relevance of the channels selected for stimulation is described. In the proposed strategy, the available channels are subdivided into groups, designated as “Selected Groups” (SG), and within each group, a specified number of active channels with the largest amplitudes are selected for stimulation. The hypothesis is that most of the spectral information that can be perceived by CI listeners is conveyed by taking the highest filter band outputs within a stimulation area represented by a group of neighboring channels. Two experiments were conducted in subjects with MED-EL implant systems, measuring recognition of sentences in speech-shaped noise. In experiment 1, the SG group size was varied from two to four while selecting one active channel per group and keeping the pulse phase durations constant. Results showed no significant difference in sentence recognition between continuous interleaved sampling and SG configurations up to a group size of three. In experiment 2, phase durations were doubled, using groups of two channels with one active channel each. This resulted in a reduction of pulse amplitudes by about 40%. Intelligibility of sentences in noise was unaffected, making a substantial reduction of implant supply voltages feasible. In all experiments, the stimulation frame rate was kept constant in order to avoid rate-change effects.

Using the Speech Transmission Index for predicting non-native speech intelligibility

While the Speech Transmission Index (STI) is widely applied for prediction of speech intelligibility in room acoustics and telecommunication engineering, it is unclear how to interpret STI values when non-native talkers or listeners are involved. Based on subjectively measured psychometric functions for sentence intelligibility in noise, for populations of native and non-native communicators, a correction function for the interpretation of the STI is derived. This function is applied to determine the appropriate STI ranges with qualification labels ("bad"-"excellent"), for specific populations of non-natives. The correction function is derived by relating the non-native psychometric function to the native psychometric function by a single parameter (nu). For listeners, the nu parameter is found to be highly correlated with linguistic entropy. It is shown that the proposed correction function is also valid for conditions featuring bandwidth limiting and reverberation.

Effect of speaking styles on the relevance of the Speech Transmission Index in the presence of reverberation

The Speech Transmission Index (STI) predicts speech intelligibility under various speech degrading conditions, including noise and reverberation. Measurements of sentence intelligibility in noise and reverberation, for three different talkers adopting different speaking styles, indicate that the STI is sometimes inaccurate in the presence of reverberation. For a trained talker speaking clearly, the STI predictions are accurate, but for conversational speech by untrained talkers, the effect of reverberation is underestimated. By using a wider range of modulation frequencies than prescribed for the standardized STI calculation (up to 31.5 Hz instead of 12.5 Hz), more accurate predictions are obtained for conversational speech. This can be understood by comparing the envelope spectrum between talkers and speaking styles: conversational speech tends to spread more of the energy in its envelope spectrum to higher modulation frequencies. Since higher modulation frequencies are more susceptible to reverberation, conversational speech tends to be affected more by reverberation than clear speech. To investigate the range of between-talker variations, envelope spectra were calculated for a population of 134 talkers. Upper boundaries of the STI modulation frequency range of 12.5 and 31.5 Hz seem approximately appropriate for the 5th and 95th percentile of this population.

Effects of spectral smearing on the intelligibility of sentences in noise

Cochlear hearing loss is generally associated with a variety of deficits in the ability to analyze sounds, including reduced frequency selectivity. The complex interrelationships between the different deficits have made it difficult to assess the contribution of reduced frequency selectivity to the difficulties experienced by hearing-impaired people in understanding speech. To address this question, the effects of reduced frequency selectivity on the place representation of speech stimuli in the auditory system were simulated by ‘‘smearing’’ the spectra of the stimuli, using the overlap–add method. The smearing was designed to evoke excitation patterns in a normal ear that would resemble those evoked in an impaired ear using unsmeared stimuli. Several different types of smearing were used, simulating specific degrees of broadening and asymmetry of the auditory filter. Sentences in quiet and in speech-shaped noise were smeared and presented to normally hearing listeners in intelligibility tests. The results showed that the intelligibility of speech in quiet was hardly affected by spectral smearing, even for smearing that simulated auditory filters six times broader than normal. However, the intelligibility of speech in noise was adversely affected by the smearing, especially for large degrees of smearing and at a low speech-to-noise ratio (−3 dB). Simulation of asymmetrical broadening of the lower side of the auditory filter had a greater effect than simulation of asymmetrical broadening of the upper side, suggesting that upward spread of masking may be particularly important.

The effect of amplitude modulation on the intelligibility of sentences in noise

Previous work has demonstrated that amplitude‐modulated time‐varying sinusoidal (AMTVS) replicas of natural sentences were more intelligible than simple time‐varying sinusoidal (TVS) sentences [T. D. Carrell, J. Acoust. Soc. Am. Suppl. 1 84, S158 (1988)]. More recent work indicated that this improvement was very likely based on the grouping effects of comodulation masking release [T. D. Carrell, J. Acoust. Soc. Am. Suppl. 1 86, S102 (1989)]. It was further proposed that amplitude modulation, which is often found in natural signals such as voiced speech, comodulates the components of a sound from a single source and may be used to help separate an auditory figure from its background in natural listening environments. The present experiment tested this claim directly. Eight TVS sentences and eight AMTVS sentences were presented in a background of multispeaker babble at a signal to noise ratio of + 6 dB. Listeners perceived only 36.9% of the phonemes in the unmodulated TVS sentences correctly, whereas they perceived 72.3% of the phonemes in the AMTVS sentences correctly. These results suggest that the amplitude comodulation of the components of a speech sound allow them to be grouped together and segregated from simultaneous sounds from other sources. This grouping and segregation is claimed to account for the improved intelligibility that was found.