Speech Intelligibility Measures Research Articles

The adaptive modified rhyme test

This paper discusses an application of the adaptive maximum-likelihood procedure (MLP) [Watson and Pelli, Percept. Psychophys. 33, 113–120 (1983)] to speech intelligibility measurements conducted with the modified rhyme test (MRT) [House et al., J. Acoust. Soc. Am. 37, 158–166 (1965)]. The MLP procedure allows the experimenter to preset the target intelligibility at any value and determine the corresponding speech-to-noise (S/N) ratio. This makes it an efficient tool for estimating the change in the S/N ratio required to maintain identical intelligibility in different noise conditions of the communication channel. The MLP-MRT was tested on a group of 12 listeners. The S/N ratio corresponding to intelligibilities of 50%, 70%, and 88% was determined in five test runs of 30 items each. Subsequently, the S/N ratios estimated by the adaptive procedure were verified by measuring the intelligibility scores by a nonadaptive method. Results showed that intelligibility scores obtained by the nonadaptive method differed by less than 5% from the intelligibility preset in the adaptive procedure. [Part of the work was done as a collaborative project between NRC and the Defence and Civil Institute of Environmental Medicine, during the author’s stay with the Acoustics and Signal Processing Group, NRC, Ottawa, Canada.]

Use of a loudness model for hearing aid fitting: II. Hearing aids with multi-channel compression.

A model for predicting loudness for people with cochlear hearing loss was applied to the problem of the initial fitting of a multi-channel compression hearing aid. The fitting was based on two constraints: (1) The specific loudness pattern evoked by speech of a moderate level (65 dB SPL) should be reasonably flat (equal loudness per critical band), and the overall loudness should be similar to that evoked in a normal listener by 65-dB speech (about 23 sones for binaural listening); (2) Speech with an overall level of 45 dB SPL should just be audible in all frequency bands from 500 Hz up to about 4 kHz, provided that this does not require compression ratios exceeding about 3. These two constraints were used to determine initial values for the gain, compression ratio and compression threshold in each channel of a multi-channel compression system. This initial fitting was based entirely on audiometric thresholds; it does not require suprathreshold loudness measures. The fitting method was evaluated using an experimental fast-acting four-channel compression system. The initial fitting was followed by an adaptive procedure to 'fine tune' the fitting, and the aids were then used in everyday life. Performance was evaluated by use of questionnaires and by measures of speech intelligibility. Although the fine tuning resulted in modest changes in the fitting parameters for some subjects, on average the frequency response shapes and compression ratios were similar before and after the fine tuning. The fittings led to satisfactory loudness impressions in everyday life and to high speech intelligibility over a wide range of levels. It was concluded that the initial fitting method gives reasonable starting values for the fine tuning.

Speech intelligibility assessment in a helium environment. II. The speech intelligibility index.

The Speech Intelligibility Index (SII) was measured for Navy divers participating in two saturation deep dives and for a group of nondivers to test different communication systems and their components. These SIIs were validated using the Speech Perception in Noise (SPIN) test and the Griffiths version of the Modified Rhyme Test (GMRT). Our goal was to determine if either of these assessments was sensitive enough to provide an objective measure of speech intelligibility when speech was processed through different helmets and helium speech unscramblers (HSUs). Results indicated that SII values and percent intelligibility decreased incrementally as background noise level increased. SIIs were very reliable across the different groups of subjects indicating that the SII was a strong measurement for predicting speech intelligibility to compare linear system components such as helmets. The SII was not useful in measuring intelligibility through nonlinear devices such as HSUs. The speech intelligibility scores on the GMRT and SPIN tests were useful when the system component being compared had a large measurable difference, such as in helmet type. However, when the differences were more subtle, such as differences in HSUs, neither the SPIN nor the GMRT appeared sensitive enough to make such distinctions. These results have theoretical as well as practical value for measuring the quality and intelligibility of helium speech enhancement systems.

An evaluation of the diagnostic rhyme test

The intelligibility of a speech output device is an important predictor of user acceptability. The Diagnostic Rhyme Test (DRT) is an ANSI standard for measuring speech intelligibility (ANSI S3.2-1989). In the DRT, respondents hear a word and choose its equivalent from two visually presented words. The two words differ only in their initial (e.g., veal-feel), and the two consonants differ only in a single distinctive acousticphonetic feature (e.g., voicing). To define “distinctive feature”, the DRT uses a minimal distinctive feature system, loosely based on the work of Jakobson et al. (1963) and Miller and Nicely (1955). These studies carefully analyzed natural speech errors in various noise environments. Whether or not these studies can be freely applied to alternative forced-choice tests of coded or synthesized speech is an empirical issue. In the present study, the results of a Consonant Identification (CI) task were compared to a previously conducted DRT using the same coding algorithms. The CI data indicated that the low-bit-rate coded speech yielded significantly more multifeature confusions then the uncoded speech. Moreover, the multifeature confusions could not be easily predicted from the single-feature confusions. A fundamental assumption of the DRT is that speech errors are adequately diagnosed by testing single-feature confusions. The results of the present study contradict that assumption. In conclusion, we argue that the application of the DRT (and more generally, any closed-response choice procedure) to coded or synthesized speech is questionable.

Assessment of speech intelligibility in large auditoria case study: kuwait state mosque

Extensive sound measurements have been conducted to assess the speech intelligibility within the main prayer hall (MPH) of the Kuwait State Mosque. Traditionally acoustical absorbing materials are used to enhance a poor speech intelligibility in a large auditorium. In this paper, however, a totally electronic approach is adopted to enhance the speech articulation in the MPH. Adaptive equalizers, as well as other advanced techniques have been adopted to enhance the overall response. The layout of the sound reinforcement system and the loudspeakers’ orientations were modified in order to minimize perturbations usually caused by reflections of sound waves and other interfering effects. These solutions have revealed a substantial improvement in speech intelligibility over the original environment of the Mosque. This paper describes the measurements of speech intelligibility in MPH, the problems associated with the original layout of the sound reinforcement system, and the effect of the relevant modifications on the improvement of the speech intelligibility in MPH.

Comparisons of measures of speech intelligibility in rooms

The speech transmission index, useful-to-detrimental ratios, and percent articulation losses are three quite different types of measures of speech intelligibility in rooms. They each combine a measure of the speech-to-noise ratio and a room acoustics measure to better relate to speech intelligibility. Variations of all three types of measures were calculated from 91 impulse responses obtained from a wide range of acoustical conditions, and for different speech-to-noise ratios. Several broadband forms of these measures are shown to be reasonably well related to each other. The results are used to derive conversion equations between the three types of measures.

Augmentative and alternative communication decision making for children with severely unintelligible speech

Decision making in augmentative and alternative communication (AAC) is particularly difficult for children who have some functional speech. This is in part because there is no measure of speech intelligibility in children that is sensitive to partner familiarity and semantic context, factors that have been shown to greatly impact the comprehensibility of severely impaired speech. This paper describes the development and clinical use of a measure of functional speech comprehensibility for children. The author presents four case studies of children with severe speech impairments who were referred for AAC evaluation. These case studies demonstrate the role of such a measurement tool in AAC decision making for the following purposes: (1) to resolve team conflicts about speech versus alternative strategies, (2) to demonstrate the benefit of speech supplementation strategies for some children for whom AAC equipment may be inappropriate at this time, and (3) to demonstrate the limitations of speech supplementation for some children who will need alternative approaches as they grow more independent.

A source impedance compensated artificial voice for mask speech intelligibility testing: Thevenin model.

This paper describes the development of a Thevenin model of the human voice for mask speech intelligibility testing. The following steps were used to develop the Thevenin model: (1) the long-term average speech spectrum of ten male speakers was measured; (2) the corresponding average vocal tract shape was estimated using linear prediction; (3) a Thevenin model of the voice was determined based on the vocal tract shape. Electroacoustic analogs are used to predict the mask speech spectra produced by the Thevenin model and a mouth simulator consisting of a loudspeaker equalized for free-air speech. Comparing the two spectra allows us to examine how the interaction between the driving point impedance of the voice and the load impedance of a mask affects objective speech intelligibility measurements. The Thevenin model is used in a voice simulator that is described in a future paper.

Fitting Hearing Aids to Individual Loudness-Perception Measures

The purpose of this study was to compare the prescribed gain, compression ratios, compression thresholds, and the relative predicted speech intelligibility (Speech Intelligibility Index [SII], American National Standards Institute 3.79, proposed) provided by four strategies proposed for selecting hearing aid parameters for low-threshold compression hearing instruments and by a traditional threshold-based hearing aid fitting procedure. The strategies used were Desired Sensation Level Input/Output (DSLTM[i/o]; Cornelisse, Seewald, & Jamieson, 1994), Visual Input-Output Locator Algorithm (VIOLA; Cox, 1994), FIG6 strategy (Killion, Reference Note 2), Ricketts and Bentler strategy (RAB), and a threshold-based hearing aid fitting procedure (National Acoustics Laboratories-Revised [NAL-R]; Byrne & Dillion, 1986). These new strategies have been suggested as alternatives to threshold-based strategies, which do not provide the varying amounts of target gain, as a function of input level, necessary to fit low-threshold compression hearing aids. The electroacoustic prescriptions and the predicted speech intelligibility were calculated across all five fitting strategies for 20 subjects. The threshold and loudness growth information used for each fitting was reported previously (Ricketts & Bentler, in press). Comparison across prescriptions revealed that the NAL-R strategy (due to the linear gain provided) prescribed the least gain for low-level inputs and the greatest gain for high-level inputs. Gain comparisons across fitting by loudness (FBL) strategies revealed a more shallow frequency response slope for strategies that require individual measures of loudness growth (RAB, VIOLA) in comparison with strategies that assumed average data (FIG6, DSLTM[i/o]). SII results revealed greater predicted speech intelligibility for the FIG6 and the DSLTM[i/o] compared with the NAL-R, RAB, and VIOLA. These differences were most apparent in noise backgrounds and least evident when loudness differences were minimized. It appears that differences in SII scores across the FBL fitting strategies are due, in part, to differences in the loudness of the output signal. It is assumed that differences in high-frequency shaping may also be a factor. These data do not appear to support the use of additional clinical time to obtain individual loudness growth measures. However, due to the fact that SII results are based on average performance, it is difficult to predict whether differences across these fitting strategies would be realized in actual measures of speech intelligibility or sound quality on an individual basis.

The Preschool Speech Intelligibility Measure

Documenting changes in speech intelligibility across time is an important but difficult task for speech-language pathologists. This study reports on the development and initial testing of the Preschool Speech Intelligibility Measure (PSIM), a single-word, multiple-choice intelligibility measure. The PSIM is adapted from the Assessment of Intelligibility of Dysarthric Speech (Yorkston & Beukelman, 1981) and is designed to plot changes in children's speech intelligibility across time. This instrument is offered as an addition to the existing array of available speech intelligibility measures.

Field trials of a portable prototype digital hearing aid

A battery‐operated digital processor connected to microphones and receivers located in left and right ear modules was built and used in a hearing aid field trial. Eight hearing impaired individuals with moderate to moderately severe hearing losses served as subjects. All subjects had symmetric hearing losses and were experienced binaural hearing aid users. Four binaural hearing aid algorithms were programmed into the processor for evaluation in the field trial. The algorithms all equalized the magnitude and phase insertion effects of the ear modules, but differed in their gain prescription. Two prescriptions based on the Articulation Index (AI), one on NAL‐R, and a control prescription were evaluated in the two week field trial. Subjects rated each algorithm in seven categories. Objective measures of speech intelligibility in noise, including measures of binaural directional hearing, were taken before and after the field trial. Intelligibility and directional hearing was best with the AI prescriptions, although these prescriptions did not receive the highest subjective ratings. Details of the binaural algorithms and fittings, as well as objective and subjective measures of their benefit will be reported.

Recruitment compensation based on a simplified loudness estimation model

To compensate for the frequency-dependent, nonlinear component of the hearing loss (recruitment), multichannel dynamic compression yields advantages over broadband compression. However, two unsolved problems remain: Due to the nonlinear additivity of perceived loudness across critical bands, the compression characteristic cannot be fitted correctly for narrow-band and broadband signals likewise. Furthermore, independent compression in different frequency bands introduces artifacts, which may limit the benefit of the algorithms. To overcome the problems, a 24-channel compression algorithm was developed and tested using critical bandwidths and time constants of ∼10 ms. The compression characteristic for broadband signals of arbitrary spectral shape is calculated by a loudness estimation model based on data from normal listeners. The algorithm should provide correct gain factors for narrow-band signals as well as for broadband signals. Furthermore, it introduces a psychoacoustically motivated interaction between frequency bands that reduces the processing artifacts. The complete algorithm was implemented in real time on a multi-signal processor setup and tested with sensorineurally impaired patients. First results of speech intelligibility measurements in quiet show the capability of the algorithm to compensate for the ‘‘distortion component’’ of the hearing loss. Further results of speech quality assessments and speech intelligibility tests in noise are presented.

Binaural hearing aid algorithms for noisy and reverberant environments

In noisy and reverberant acoustical environments the human hearing system is capable of concentrating on a desired speaker, while attenuating undesired sounds like noise or reverberation. This so called ‘‘cocktail-party processing’’ is reduced in hearing-impaired persons. New hearing aid strategies try to imitate binaural processing principles to suppress interfering noise or reverberation in order to yield maximum speech intelligibility for the impaired listener. Two algorithms are described that reduce unwanted signal energies by attenuating sound that arises from the sides (lateral suppression) and that remove reverberation and short reflections. They were combined and implemented in real time on a multiprocessor setup to allow for adaptive adjustment of parameters and an optimal fit to the individual hearing loss. The noise reduction performance of the algorithms was tested with normal- and hearing-impaired subjects. Subjective speech quality and speech intelligibility assessments as well as speech intelligibility measurements show that both groups benefit significantly from the preprocessing strategies in anechoic and reverberant acoustical situations with interfering speakers.

Assessing intelligibility

Accurate measurement of speech intelligibility is a critical component for the assessment of children's disordered phonologies. The degree of intelligibility has important implications for planning and implementing intervention and also for evaluating its effectiveness. Limited normative data are av

Effects of listeners' experience on two measures of intelligibility.

10 speech-language pathologists with extensive experience in judging speakers' intelligibility and 10 control subjects with no such previous experience provided written identification and magnitude-estimation scaling judgments of the intelligibility of nine audiotaped speech samples. Analysis indicated no significant main effect for experience on either the written identification or the magnitude-estimation scaling tasks. Implications for the continued use of magnitude-estimation scaling as a measure of speech intelligibility are discussed.

EFFECTS OF LISTENERS' EXPERIENCE ON TWO MEASURES OF INTELLIGIBILITY

10 speech-language pathologists with extensive experience in judging speakers' intelligibility and 10 control subjects with no such previous experience provided written identification and magnitude-estimation scaling judgments of the intelligibility of nine audiotaped speech samples. Analysis indicated no significant main effect for experience on either the written identification or the magnitude-estimation scaling tasks. Implications for the continued use of magnitude-estimation scaling as a measure of speech intelligibility are discussed.

Beyond intelligibility testing: Development of an information theory based voice communication effectiveness test.

Speech intelligibility has been a popular measure of voice communication effectiveness. However, intelligibility measures do not account for some critical parameters in voice communication effectiveness. Task complexity, task criticality, language complexity, and time, all affect voice communication effectiveness but are not factors in speech intelligibility measures. Information theory gives the conceptual and mathematical basis for quantifying these parameters in terms of entropy, mutual information, channel rate, and channel capacity. The Voice Communications Effectiveness Test (VCET) has been developed to directly measure channel capacity of a communication ‘‘system’’ by varying these critical information theory parameters. VCET consists of a two-way interactive communication task. Scoring is based on correct responses and time. The basis for the development of VCET, its structure, application, and preliminary performance data will be presented.

Measurement and predictions of hearing handicap using an additive noise model.

The ability to accurately, quickly, and reliably measure speech intelligibility in noise is an essential requirement for assessing the benefit of hearing aids to speech communication in noisy environments. Measurement error caused by variability of speech materials, content, speaking styles, talkers, and spectral characteristics of noise maskers often make accurate intelligibility measurements impractical. This paper will review measurements of speech intelligibility as applied to the hearing impaired. Emphasis will be placed on Dutch sentence tests developed by Plomp [e.g., R. Plomp, J. Acoust. Soc. Am. 63, 533–549 (1978)] and on recent English sentence tests we have developed, which control many of the sources of measurement error. Test development procedures will be described, and a summary of initial results will be used to demonstrate the reliability of the tests for both group and individual assessment of intelligibility in noise.

Comparison of Objective and Subjective Measures of Speech Intelligibility in Elderly Hearing-Impaired Listeners

Three experiments were performed to evaluate the use of subjective intelligibility estimations as a method for measuring hearing aid benefit. Subjective and objective speech intelligibility scores were compared for young normal-hearing and elderly hearing-impaired listeners. Objective intelligibility scores were obtained using the Connected Speech Test (CST). This test consists of conversationally produced passages of speech that the listener repeats sentence by sentence. To provide subjective intelligibility scores, listeners estimated the percentage that they understood of each CST passage. Comparison of the two types of scores revealed that they were closely related in both groups of subjects (r = .82-.92). Although the two types of scores were essentially equal for normal-hearing subjects, the hearing-impaired listeners tended to produce subjective estimations of intelligibility that were significantly lower than their objective scores. Manipulation of visual cues and amplification, in an attempt to influence the hearing-impaired listeners' expectation of understanding speech, had no effect on the subjective-objective score differential. The difference between subjective and objective scores in the hearing-impaired group was not related to audiometric variables such as speech reception threshold, audiogram, or duration of hearing loss. It was concluded that comparative hearing aid evaluations using subjective intelligibility estimates would usually produce the same relative outcome as evaluations using the objective intelligibility measurement procedure. However, scores obtained with the objective procedure had smaller critical differences. Thus, when both types of scores are based on the same number of passages, the objective measurement procedure would be the more sensitive to differences among hearing aids.

Computerized Measurement of Speech Intelligibility I.: Development of System and Procedures

An experimental system for the measurement of speech intelligibility has been developed. It uses a Personal Computer (PC), together with appropriate software to handle playback of test words in carrier phrases, presented in a 'closed response' test condition. Information about the intelligibility, based on the correct responses and the confusions, is immediately available due to simultaneous collecting and sorting of subjects' responses. The system works satisfactorily and reliably and has been well received by experimenters as well as by adult test subjects in the age range 18-70 years. From a new Danish standard speech material for audiological purposes, a Multiple Choice speech intelligibility test has been devised. The test is called 4AFC (Four Alternative Forced Choice) and is based on monosyllabic words with consonant confusions. Normative data for the 4AFC test, obtained with the computerized system, are presented in an accompanying paper.