Search

Measurement procedures using broadband noise stimuli for testing hearing aids have recently been standardized in an attempt to ensure that test results more closely reflect the ‘‘real‐world’’ performance of hearing aids. Although a number of researchers have employed coherence measures made with broadband noise stimuli to characterize the broadband distortion of hearing aids, broadband distortion measurement methods have not yet been standardized. In this paper, it is demonstrated through simulations and measurements on hearing aids that similar results are obtained when coherence is measured using maximum length sequence (MLS) and conventional unbiased statistical‐based methods. The calculation of single‐ and dual‐channel MLS coherence are also formalized. Signal‐to‐distortion ratio (SDR) measurements for two automatic signal processing hearing aids are presented.

There is a growing trend for hearing aids to incorporate wide dynamic range compression. The input/output (I/O) hearing aid formula, presented in this report, is a general frequency‐specific mathematical approach which describes the relationship between the input level of a signal delivered to a hearing aid and the output level produced by the hearing aid. The I/O formula relates basic psychoacoustic parameters, including hearing threshold level and uncomfortable listening level, to the electroacoustic characteristics of hearing aids. The main design goal of the I/O formula was to fit the acoustic region corresponding to the ‘‘extended’’ normal auditory dynamic range into the hearing‐impaired individual’s residual auditory dynamic range. The I/O approach can be used to fit hearing aids utilizing linear gain, linear compression or curvilinear compression to a hearing‐impaired individual’s residual auditory area.

CSRE (The Canadian Speech Research Environment) is a comprehensive, microcomputer‐based system designed to support speech research using IBM/AT‐compatible micro‐computers. CSRE provides a powerful, low‐cost facility in support of speech research, using mass‐produced and widely available hardware. Functions include speech capture, editing, and replay; several alternative spectral analysis procedures, with color and surface/3‐D displays; parameter extraction/tracking and tools to automate measurement and support data logging; alternative pitch‐extraction systems; parametric speech (KLATT80) and nonspeech acoustic synthesis, with a variety of supporting productivity tools; and a comprehensive experiment generator/controller, to support behavioral testing using a variety of common testing protocols. The CSRE project is nonprofit, and relies on the cooperation of researchers at a number of institution. Version 3.0 of CSRE has been used since 1989 by researchers in 12 countries. Version 4.0 offers a wider range of functions, runs faster, uses higher resolution displays, and supports additional hardware systems, including digital signal processing boards.

Typical measurements of electroacoustic performance of hearing aids include frequency response, compression ratio, threshold and time constants, equivalent input noise, and total harmonic distortion. These measurements employ artificial test signals and do not relate well to perceptual indices of hearing aid performance. Speech‐based electroacoustic measures provide means to quantify the real world performance of hearing aids and have been shown to correlate better with perceptual data. This paper investigates the application of system identification paradigm for deriving the speech‐based measures, where the hearing aid is modeled as a linear time‐varying system and its response to speech stimuli is predicted using a linear adaptive filter. The performance of three adaptive filtering algorithms, viz. the Least Mean Square (LMS), Normalized LMS, and the Affine Projection Algorithm (APA) was investigated using simulated and real digital hearing aids. In particular, the convergence and tracking behavior of these algorithms in modeling compression hearing aids was thoroughly investigated for a range of compression ratio and threshold parameters, and attack and release time constants. Our results show that the NLMS and APA algorithms are capable of modeling digital hearing aids under a variety of compression conditions, and are suitable for deriving speech‐based metrics of hearing aid performance.

This study investigated the effects of perceptual and production training on the abilities of adult native speakers of Spanish to identify and produce the English vowels /i, I, e, e, ae/. Testing was performed prior to and following an average of 12 h of perceptual training in a category inclusion task, then again following an average of 7.5 h of production training involving visual feedback of vowel formant (F1 and F2) values. A lagged control group of participants who were delayed in starting training showed no improvement in perception and production skills during the control period but changed equivalently during training. The mean improvement in perceptual identification accuracy for /i, I, e, e/ was 17%. The mean improvement in the intelligibility of participants productions of /I, e, ae/ following training was 11%. For both perception and production, most improvement occurred during perceptual training.

This paper considers an acoustic basis for speech intelligibility and evaluates various acoustically based speech intelligibility prediction algorithms. Earlier research indicates that speech intelligibility does not require preservation of spectral and temporal fine‐structure, but is highly dependent on the preservation of the amplitude component of the modulation spectrum [R. Drullman, J. Acoust. Soc. Am. 97, 585–592 (1995)]. This study assessed the importance of the phase component of the modulation spectrum using a 21‐alternative forced‐choice consonant perception test. Temporal and spectral fine‐structure were removed by modulating a white noise carrier with 50 Hz low‐pass filtered speech amplitude envelopes in 4, 8, or 24 discrete bands. Modulation spectrum phase was distorted by imposing a random delay in each discrete band. Behavioral results are discussed in light of intelligibility predictions generated by the articulation index [N. R. French and J. C. Steinberg, J. Acoust. Soc. Am. 19, 90–119 (...

Subjective assessments of the severity of hearing aid distortion were obtained using both rating scale and paired comparison measures, with sentence materials. Data were collected under 30 hearing aid conditions, representing a broad range of real hearing aid distortions. These data were related to a range of objective (engineering) measures of hearing aid distortion, including techniques based on pure tones (THD, IMD), measures using broadband noise stimuli having certain speech‐like characteristics, and measures derived from real speech. Data from both normal and hearing‐impaired subjects were stable and orderly, with good agreement between datasets collected using paired comparison and rating scale methods. Correlations with behavioral data were significantly better for certain speech‐based objective measures than electroacoustic measures based on noise or pure‐tone signals. [Work supported by the Natural Sciences and Engineering Research Council of Canada.]

The electroacoustic properties of many modern hearing aids change as a function of the input stimuli. One consequence is that electroacoustic measurements obtained using commercial hearing aid test systems may differ from those made with important, ‘‘real‐life’’ signals. As speech is the most important real‐life signal for most hearing aid users, tests using speech stimuli have been proposed. This paper describes one approach to such testing, using spectrograms to display speech signals in a three‐dimensional representation. Dual spectrogram displays prepared using the computerized speech research environment (CSRE) software allow speech signals at the input and output of a hearing aid to be compared. Such spectrograms are computed for a set of important phonemes extracted from running speech. The results of these analyses are evaluated for individual hearing aid users using a display which includes the user’s thresholds and loudness discomfort levels (LDLs), together with the target amplification levels ...

A mathematical model is described for mechano‐electro‐chemical transduction in hair cells. The model relates the displacement of stereocilia to the intensity of firing of fibers attached to each hair cell through a cascade of four stages. The gating stage relates hair bundle displacement to the change in conductivity of the hair cell membrane. It is described by a mechanical version of the Hodgkin–Huxley model. The electric stage relates the change in conductivity to the receptor potential through a hair cell electrical network of the Davis type. The calcium kinetics stage relates free calcium ion (Ca2+) concentration in hair cells to the receptor potential. It is assumed there are two states of calcium ions: free and bound (by buffers). The synaptic transmission stage relates free calcium concentration to the intensity of firing as described by the Hill equation. Using relevant parameters, the model simulates the behavior of low‐ and high‐spontaneous firing rate fibers, and is being implemented as a comp...

An automated dual‐channel maximum‐length sequence (MLS) test system for the electroacoustic characterization of hearing aids has been developed. This test system applies a speech‐shaped MLS acoustically to a hearing aid and measures the electro‐acoustic frequency response. This method provides results that compare favorably to those obtained using the method employed by Kates [J. M. Kates, J. Rehab. Res. Develop. 27, 255–278 (1990)] and the ANSI standard method [ANSI S3.42 (1992)]. Test results show that MLS‐based testing is significantly faster than noise‐based testing. Two signal‐biased MLS‐based testing methods have also been developed. These methods apply a bias signal to force the hearing aid into a mode of operation where the frequency response, with adaptive or subtractive bias signal cancellation, is measured using a low‐level, speech‐shaped MLS. This method has proved valuable for the characterization of automatic signal processing hearing aids. [Work supported by ORTC and OMH.]

The experiment control system is a package of computer programs that facilitate the generation and conduct of identification and discrimination experiments using speech and other complex acoustical signals. The system permits microcomputers to be used to control experimental equipment, present signals to listeners, and record and analyze responses, without preparing individual computer programs for each experiment. The package contains three general programs. The GENERATE program is used to define experimental parameters via full‐screen data entry forms. The core of the program is an entry form that allows one to specify four basic parts of an experimental trial: a warning signal, one or more stimulus presentations, the response(s) to be recorded, and any feedback to be provided after the response. Within each of these categories, a number of alternatives are permitted. The CONTROL program applies the parameters and data defined in GENERATE to run the experimental session. Typically, possible responses are displayed using high‐resolution graphics on a computer video screen and subjects respond using a computer mouse. The SCORE program analyzes output from CONTROL and displays the results as response matrices with basic statistics such as means and standard deviations, as well as row and column sums. Output is provided as a screen display and print‐ready file. ECoS is being made available to researchers as part of Version 4.0 of CSRE (the Canadian Speech Research Environment).

A common complaint of hearing aid wearers is that they have difficulty when listening in a background of noise. The ability of a digital adaptive noise reduction filter to reduce background noise without degrading speech perception has been evaluated. Each of four normal‐hearing and six hearing‐impaired subjects completed two behavioral tests of speech perception—the speech reception threshold (SRT) and the modified distinctive features differences test [DFD(m)]—and one subjective measure of listener preference (paired comparisons). Each test was performed in three types of background noise presented at a level of 65 dB: white noise, low‐pass filtered at 8.0 kHz; white noise, low‐pass filtered at 1 kHz; and multitalker speech babble. For hearing‐impaired subjects, the speech‐plus‐noise signal was filtered and amplified to provide the NAL real‐ear insertion gain. Results indicated that all subjects preferred the noise‐reduction‐processed speech‐plus‐noise signal over the unprocessed signal at positive sign...

This paper compares four alternative procedures for estimating the acoustical characteristics of speech and other complex signals: the fast Fourier transform (FFT), two autoregressive techniques (the autocorrelation and modified covariance methods), and the Cone–Kernel method. All four techniques are now widely available to researchers in Version 4.0 of the CSRE (Canadian Speech Research Environment) software system. FFT spectral estimates are characterized by many known tradeoffs: good frequency resolution is obtained at the expense of good time resolution, and vice versa. Autoregressive methods overcome some of the inherent limitations of the FFT method, but still use a quasi‐stationary approach to analyze nonstationary signals such as speech. The Cone–Kernel method offers the opportunity for both good frequency and good time resolution, and cancels many of the cross terms that have made it difficult to interpret previous approaches based on the Wigner distribution. Examples of applying the various analysis procedures to signals will be shown, including situations where the Cone–Kernel approach facilitates the interpretation of certain details of the speech signal.

Cue deletion was used to examine the contribution of various acoustic cues to the identification of the English stop consonants (/b/, /d/, /g/, /p/, /t/, /k/). Stimuli were based on tokens spoken by four talkers (two men and two women) in a syllable‐medial context (/(inverted vee) Cil /). Cues within five regions centered on the burst were examined: (1) the transition from the preceding vowel into consonant closure, (2) the closure interval, (3) the consonant release, (4) the initial transition from release to the following vowel, and (5) the final transition to the vowel. Each region either remained at the full normal amplitude or was attenuated fully (i.e., effectively replaced by silence). Four listeners who had normal hearing identified the 768 signals constructed from each of the possible combinations of attenuated and full‐amplitude regions (i.e., 32 combinations ×6 stimuli ×4 speakers). Analyses of these identification responses demonstrates the relative importance of each region, and of each combination of regions, for consonant identification. Acoustic measures explain some aspects of the confusion errors which occurred.

Subjects discriminated a "standard" pair of tone bursts (T1, T2) from a "comparison" pair (T1 + delta t, T2 + delta f), containing increments in the duration delta t of the first burst and/or the frequency delta f of the second burst. The threshold (d' = 2.0) for delta t was measured as a function of delta f, and the threshold for delta f as a function of delta t. The integration of increments in duration and frequency was studied as a function of the spectral and temporal separation between T1 and T2. A trade-off between the values of delta t and delta f required for d' = 2.0 performance was observed. This integration takes place when delta t, delta f occur simultaneously in the same spectral region, and when they occur separated by up to 120 ms, or by up to a full octave. The efficiency of integration was similar for all conditions of temporal and spectral separation studied, because the discriminability of delta t and of delta f is also nearly uniform across experimental conditions. The results from all experimental conditions are adequately described by a vector summation model derived from TSD. In a subsidiary experiment, subjects categorized pure tones varying in duration and frequency as "high" or "low" in pitch and "long" or "short" in duration. It was found that combined variations in duration and frequency result in essentially independent perceptual processes, although pitch has a small effect upon the perceived duration. It is concluded that spectral-temporal integration is a general ability operating in a variety of stimulus conditions.(ABSTRACT TRUNCATED AT 250 WORDS)

A perceptual fading technique [cf. Jamieson and Morosan, Percept. Psychophys. 40, 205–215 (1986)] was used to train two types of listeners to perceive new fricative contrasts. One group consisted of young adult francophones who had difficulty producing and perceiving the voiced/ voiceless “th” distinction in English. With this group, training using the fading technique improved identification accuracy both for the specific, synthetic targets used in training, and for multiple, natural tokens of these sounds, spoken by different individuals, both men and women. However, this learning was quite specific, as performance was poor when listeners were tested with the target sounds in new word positions, or when /d/ tokens were used in the distractor set. The second group of listeners were young children who consistently mispronounced at least one English fricative sound, and had been diagnosed as having a functional articulation disorder. A sizeable proportion of these children, all of whom had been selected on...

When listeners are asked to discriminate changes in the duration of a brief frequency glide which is followed immediately by a longer, steady‐state, pure tone, they show better performance when the duration of the glide is in the range 40–60 ms than when it is longer or shorter. This effect could result from the interaction of two psychoacoustic functions—the first, a discriminability function, which follows Weber's Law and the second, a masking function. Three experiments were undertaken to test this explanation. In experiment 1, the glide began at 400 Hz, ended at 700 Hz, and was immediately followed by a steady‐state pure tone of equivalent amplitude, also at 700 Hz. Overall stimulus duration was fixed at 200 ms, while glide and steady‐state values varied in 10‐ms intervals, as follows: 10,190; 20,180; 30,170; 40,160; 50,150; 60,140; 70,130; 80,120; 90,110. This experiment replicated previous results [D. Jamieson and E. Slawinska, J. Acoust. Soc. Am. Suppl. 1 74, S67 (1983)] in finding the discriminability (− ln η) was a maximum and bias (ln β) a minimum, when glide duration was in the interval 40–60 ms. Stimuli for experiment 2 were identical to those for experiment 1, except that the amplitude of the steady‐state pure tone was reduced 10 dB. Stimuli for experiment 3 were identical to those for experiment 1 except that the overall duration of the stimulus was fixed at 10 ms. Thus, glide and steady‐state pure‐tone components had the following durations in ms: 10,90; 20,80; 30,70; 40,60; 50,50; 60,40; 70,30; 80,20; 90,110. Results obtained in experiments 2 and 3 are consistent with the hypothesis that the effect seen in experiment 1 is due, in part, to the masking of the glide stimulus by the following steady‐state pure tone. In both cases, the point of maximum discrimination (− ln η) shifted to shorter glide durations (20–40 ms) than was found in experiment 1.

Breath sounds in prematurely born infants are thought to be less “vesicular” than in term infants. This difference in sound quality in newborns of various gestational ages can be explained by different transmission characteristics of lung parenchyma. Chest sounds of 27 infants, of gestational ages between 30 and 40 weeks, were recorded at different chest locations. The results show that there is a strong correlation between chest sounds of normal babies and the location of the recording microphone on the infant's chest. The spectra of inspiration of normal term babies recorded at the level of the lower posterior chest, show dominating amplitudes in the lower frequencies (350 Hz). Sounds recorded at the upper posterior chest, show spectra with enhanced amplitudes in a broader range of higher frequencies (300–900 Hz). The studies also suggest the existence of a marked difference between breathing patterns of prematurely and term born infants. Sounds of premature babies tend to have a lower overall intensity, and spectra with dominant amplitudes in a narrower bandwidth. Moreover, dominant amplitudes are shifted toward higher frequencies as compared to sounds recorded in term infants.

Accurate identification of front‐back vowels depends on the ability to locate the peak of F2. Since hearing impairment reduces frequency resolution, sensorineural hearing impaired listeners often have difficulty with certain vowel distinctions. Espinoza‐Varas, Jamieson, and Wahn [J. Acoust. Soc. Am. Suppl. 1 76, S80 (1984)] showed that (a) for normal listeners, equal changes in F2 frequency in critical‐band units produce equal changes in the judged “goodness” of vowels on front‐back vowel continua, cued by changes in F2; and (b) sensorineural impaired listeners are less sensitive to these F2 changes and less accurate in their identification of vowels. The present paper continues this work, examining whether vowel sounds can be altered to reduce masking between adjacent spectral frequencies, while preserving vowel quality. Following Assmann [J. Acoust. Soc. Am. Suppl. 1 76, S81 (1984)] we examined identification for synthesized vowels with normal versus narrow formant bandwidths. We also examined vowel ide...

Researchers now enjoy an increasingly wide selection of alternative procedures with which to characterize the time × frequency × amplitude variation in a speech waveform. Spectral analyses using a filter‐bank technique, the digital Fourier transform (DFT), and a variety of linear predictive coding (LPC) procedures are now widely available. Approaches using an autoregressive (AR), moving average (MA), auto‐regressive moving average (ARMA), or maximum entropy (ME) procedure are already widely available, or are becoming available. The present paper reviews the assumptions of, and the interrelations between, alternative methods, and outlines the relative advantages of the various approaches in providing information about the speech waveform. Finally, the human factors aspects of alternative spectral displays are discussed.

We examined a possible psychoacoustic basis for the relatively invariant duration over which the rapid formant frequency transitions cueing place of articulation occur in initial‐position stop consonants such as /ba/. Our stimuli consisted of a 200‐ms duration compound of a pure‐tone frequency glide followed by a steady‐state pure‐tone stimulus. Glides covered a fixed range of frequencies (400–700 Hz; 1500–2200 Hz; or both together) over a variable interval of time (10–90 ms, in 10‐ms step). Three experiments examined discrimination performance in an AX paradigm; on any trial X = A or X = (A ± one step in a stimulus series). In each experiment, discriminability (−lnμ) tended to achieve a maximum, and bias (lnβ) tended to achieve a minimum, when the transition occurred over 50 ms. These results support and extend earlier reports [A. Nabelek and I. Hirsch, J. Acoust. Soc. Am. 45, 1510–1519 (1969)] and appear to reflect a special perceptual sensitivity to frequency changes which occur within the 40–60‐ms int...