Confusion Patterns Research Articles

Recognition of emotion from facial expression in mental handicap.

When mildly and severely mentally retarded people saw photographs showing expressions of basic emotions their overall performance was found to be correlated with intelligence. Certain specific patterns of confusion were determined, clinically the most significant being an inability to cope with high intensity emotions.

Misreadings by beginning readers of Serbo-Croatian.

Errors in reading aloud by the beginning reader have been interpreted as reflecting the difficulty and the importance of phonemic segmentation for the acquisition of reading skills. Results from previous studies on English words patterned as consonant-vowel-consonant showed: (1) more errors on vowels than on consonants; (2) more errors on word final consonants than on word initial consonants; and suggested that (3) consonant errors were based on phonetic confusions while vowel errors were not. In contrast to their English counterparts, the beginning readers of Serbo-Croatian tested in the present study committed proportionally fewer errors on their reading of vowels than of consonants but in common with their English counterparts, their reading of final consonants was more vulnerable to error than their reading of initial consonants. This pattern of errors was found for both word and pseudoword consonant-vowel-consonant structures and the pattern of vowel confusion, like the pattern of consonant confusions, was rationalized by speech-related factors. The differences between the patterns of confusions for Serbo-Croatian and for English could be due to the difference between the two orthographies in the precision with which they represent the phonology or to the fact that the vowels of English are qualitatively less distinct phonologically than the vowels of Serbo-Croatian.

Identification of tonal contrasts in Thai aphasic patients

In tone languages pitch variations (tones) serve to distinguish the lexical meanings of words. This study was conducted to examine the extent and nature of impairment in the perception of tones by aphasic patients who were monolingual speakers of Thai, a tone language which has five contrastive tones (mid, low, falling, high, rising). Six subjects participated in the study: two Broca aphasics, one transcortical motor aphasic, one conduction aphasic, one right brain-damaged nonaphasic, and one normal control. Three sets of stimuli (two real-speech, one synthetic-speech) were presented for identification, each set containing five Thai words minimally distinguished by tone. Results of the perception tests indicated that the performance of all four left brain-damaged aphasics differed significantly from that of the normal control, while the performance of the right brain-damaged nonaphasic did not. The normal performance of the right brain-damaged nonaphasic patient on this tone identification task suggests that deficits in the perception of tone exhibited by left brain-damaged patients can be attributed specifically to pathology in the language dominant hemisphere rather than to a general brain-damage effect. No difference in performance among the left brain-damaged patients could be attributed to a specific type of aphasic syndrome. The pattern of tonal confusions of the aphasics in comparison to that of normals suggests that their deficit is primarily quantitative rather than qualitative. Although two (mid, low) of the five tones accounted for a large percentage of the aphasics' errors, no uniform rank order of tones in terms of identifiability could be established across aphasic subjects, which suggests that their deficit is general to all five tones rather than selective to individual tones.

Evaluation of hearing-impaired listeners using a Nonsense-syllable Test. II. Syllable recognition and consonant confusion patterns.

Syllable recognition ability and consonant confusion patterns were evaluated for 38 listeners with mild-to-moderate sensorineural hearing loss using the closed-set Nonsense-Syllable Test (NST). Performance for these materials varies as a function of consonant voicing, the position of the consonant in the syllable, and the accompanying vowel. Scores for listeners with steeply sloping audiometric configurations were consistently poorer than those for listeners with gradually sloping or flat audiograms. Consonant confusion analyses revealed place of articulation errors to be the most frequent, regardless of the listener's audiometric configuration. Analysis of consonant confusion patterns indicates the existence of a systematic relationship between consonant confusions and audiometric configuration. The NST findings are discussed in terms of the test's potential use and are compared to the results of existing confusion analyses.

Identification of Japanese syllables by Spanish-speaking listeners.

In a previous work we have obtained evidence to the phonetic similarity between the five Japanese and Spanish vowels. A perceptual correspondence between syllables of both languages was also indicated. In the present study we have included a larger sample of Japanese syllabic sounds. Similarities and differences were determined by using identification tests under normal and distorted conditions (syllable played backwards). One hundred Japanese syllables were presented to ten Spanish-speaking listeners. They were instructed to listen to and to transcribe the sounds using their own Spanish alphabetic symbols. All five vowels were identified with high accuracy, whether presented alone (96%), with another vowel (97.5%), or paired with consonants (94%). The consonants in CV syllables reached an average level of identification of 89% and CVV syllables of 72%. When presented under distorted conditions, patterns of identification and patterns of confusion for each one of the sounds, common to both systems, were similar. Vowels remained unchanged. With exception of stops [p][t][k] and [d] and fricative [3], consonantal sounds were also reversible. It is concluded that a considerable number of Japanese syllabic sounds have a perceptual equivalent in the Spanish syllabic system. The acoustical distribution within the temporal syllabic segment seems to be a relevant factor in the perceptual similarity.

Computer-Baesed Recognition of Perceptual Patterns and Learning Styles in Rhythmic Dictation Exercises

During the 1978-79 academic year, sixteen freshman music majors participated in an experiment in which their response data were saved as they worked through 24 units of rhythmic dictation exercises in the University of Delaware's GUIDO® system. Analysis of the student data base showed perceptual patterns and learning styles common to exercises in both simple and compound meters. First, it was found that basic undotted, nonduplet, nontriplet notes are confused exclusively with themselves and never with dotted notes, duplets, or triplets. The same exclusive confusion pattern is seen for dotted notes, duplets, and triplets, except that they are also confused with their unmodified basic counterparts. By varying the time signatures used in the experiment, it was found that significantly more exercises are correctly answered in simple meter when a four is on the bottom, and the same pattern was found when an eight is on the bottom in compound meter. Randomly varying the pitch of the monotone stimulus between c and c2 had no effect on student achievement. As the level of difficulty increased in the 24 units, so did the average student response time and the number of times students asked for the stimulus to be replayed. The average speed at which students played the stimulus also decreased. Students who used high average dictation speeds tended to request fewer repetitions of the stimulus, as did students who used the metronome. However, neither speed of dictation, use of the metronome, nor number of repetitions had a high correlation with student achievement.

Predicting consonant confusions from acoustic analysis.

Acoustic measurements of nonsense syllables in quiet and in noise were used to predict the pattern of consonant confusions made under those conditions. Eleven sets of nonsense syllables were presented to normal-hearing subjects in quiet and at a +5 dB speech-to-noise ratio, at five speech levels. A set of acoustic characteristics of the speech stimuli were chosen for analysis and measured using digital processing techniques. Results of the recognition task revealed significant effects of consonant voicing, position and vowel context on syllable recognition. The performance-intensity function of the quiet condition rises more steeply than the function obtained in noise. The effect of noise on consonant recognition is dependent upon the manner in which the consonant is produced, and the location of maximum constriction. Differences in the absolute values of the acoustic parameters of syllable pairs were used to predict their percentage of confusion. A set of acoustic variables was isolated which was found to be the best predictor of confusion percentages. Although the sets of acoustic variables were different for various syllable types and test conditions, three variables (consonant energy, consonant spectral peaks, consonant-to-noise ratio) were used in a majority of the predictions.

Perception of amplitude‐compressed speech by impaired listeners

CV syllable identification by practiced hearing‐impaired listeners was compared for amplitude‐compressed and linearly amplified speech. The amplitude‐compression system (a 16‐channel syllabic compressor) and the linear‐amplification system were configured according to previously determined procedures [R. P. Lippmann, L. D. Braida, and N. I. Durlach, J. Acoust. Soc. Am. 62, S90(A) (1977)]. Subjects were trained for about 30 hours to identify syllables presented at random from a set of 576 stimuli representing 24 consonants, three vowels, four speakers, and two utterances. Performance after training indicated no advantage in compression for moderate hearing loss, consistent with Lippmann's results. In the case of a severe flat loss, however, a small advantage was observed for compressed speech. Some confusions in identification were specific to the type of processing, although general patterns of confusion were roughly similar for compressed and linear speech. All confusions were further analyzed in terms of articulatory features and prominent confusions were examined more closely using a discrimination paradigm. [Work supported by NIH and NSE.]

Recognition confusions among quantifiers

Subjects listened to a story containing sentences with five different quantifiers (all, many, some, a few, and none). They were then given a recognition test in which for each sentence they were asked to rate all the quantifiers as to how certain they were that each had appeared in the sentence. The degree of confusion between any two quantifiers (measured by both mean certainty ratings and first choice frequencies) declined monotonically with the separation of the two terms in a linear order. The confusion pattern was identical regardless of whether subjects were given intentional or incidental learning instructions. Successive interval scaling and the Luce biased choice model were both used to derive unidimensional scales of quantifier values that accurately described the obtained confusion matrix. A semantic feature model was also able to provide a fairly good overall account of the data, except that it had difficulty explaining the high frequency of confusions between all and many. The results raise the possibility that a major component of the memory trace for quantifiers is an essentially analog representation of magnitude.

Comment on recognition and confusion of the lowercase alphabet.

In a recent issue of this journal, Geyer (1977) presents an analysis of a full alphabet matrix for lowercase letters foveal vision. Geyer indicates previous data for the lowercase alphabet (Bouma, 1971) are inconclusive. Bouma presents two experiments which the procedures used to control average correct performance differ substantially. In one experiment, letters are presented foveally, correct performance being limited to 50% by manipulating viewing distance. The second experiment manipulated the degree of eccentric exposure along the horizontal meridian, to achieve the desired level of performance. forced-choice procedure was used the latter experiment but not the former, and the letter y was omitted from the stimulus set the nonfoveal condition. Bouma concluded on subjective grounds the foveal and nonfoveal matrices were sufficiently similar, and thus averaged the results across the foveal and nonfoveal matrices subsequent analyses. Geyer (1977) suggests the comparability of processing between foveal and nonfoveal regions, based on identification performance, must be considered an empirical question. Geyer attempts to answer this question by comparing the matrix reported by Bouma (1971) for eccentric vision with the foveal matrix presented Geyer (1977). He makes this comparison on the basis of the sum square deviation of (1) the foveal and nonfoveal matrices (2.08), and (2) the foveal matrix and a dummy matrix (1.79), uniformly composed of 0.5 along main-diagonal cells and 0.02 elsewhere (Geyer, 1977, p, 489). The use of this technique the evaluation of matrix data is not unprecedented. Townsend (1971a, 1971b) suggests the sum of squared deviations as an index of the accuracy of fit between predicted and empirical matrices. However, caution must be exercised the interpretation of this index. Geyer assumes this difference to be a significant one. The statistical basis for this assumption is not entirely clear. How large a difference is significant? Without a knowledge of the underlying distribution, such an inferential statement cannot be made. Geyer concludes the difference between sum square deviations may be regarded as evidence support of the notion that foveal and nonfoveal recognition processes differ qualitatively (p. 489). This statement is somewhat premature and may overstep the facts. No mention is made regarding the manner which the missing data from Bouma's non foveal matrix are treated. If the row for y (y was not presented) Bouma's 25 by 26 matrix is replaced by inserting .00 all of the cells, then the sum square deviation is artificially inflated. Some mention of the way which this problem was dealt with is essential to the interpretation of the comparison. Over and above the statistical question, there is perhaps a more fundamental criticism to be considered. The stimuli Geyer's study were from the Tactype Futura demi 5424 alphabet, a font without serifs. Geyer states this type style was chosen in order to minimize the influence of any particular type style idiosyncracy (p. 488). Bouma (1971) used letters of the typeface Courier 10 (IBM), a font which has pronounced serifs (p. 460). An inspection of the two type styles reveals there are considerable differences structure between corresponding letter pairs. few examples will serve to illustrate the potentialconfound inherent the comparison of different type styles. The letter A Tactype Futura demi 5424 (a) shares very few features common with the same letter Courier 10 (a) type. The similarity of Geyer's ( a) to the letter ( 0 ) is reflected the 14010 response between these two when (a) is the stimulus. Contrast with this the lack of reported by Bouma, between the same two letters (a) and (0) when (a) is the stimulus (1010). more compelling example is seen the 31010 reported by Bouma when (g) is the stimulus and (q) the response. The same occurs only 7010 of the time for Geyer's (g) and (q ). The difference structure and of itself, between these pairs, might be an important contributing factor to the differences between the foveal and nonfoveal matrices. It seems questionable whether the differences reported by Geyer (1977) can be unequivocally attributed to differences information processing between different retinal loci. It may be unreasonable to assume differences type style can be disregarded this way. Maddox, Burnette, and Gutmann (1977) present data for uppercase dot matrix letters, which indicate letter style may be an important variable altering the of confusions. Further, research presented by Fisher, Monty, and Glucksberg (1969) suggests comparison across data derived from different procedures may not be valid. The comparison made by Geyer (1977) rests on the assumption a basic pattern of confusion exists between the lowercase letters of the

Spatial frequency analysis of block letters does not predict experimental confusions.

To test the generality of spatial frequency channels models of vision, matrices of similarities between spatial frequency compositions of letters of the alphabet were generated mathematically. Methods and assumptions were chosen to reflect probable neural underpinnings of spatial frequency analysis mechanisms. These similarity matrices were tested against experimentally produced confusion matrices. The similarity matrices predicted subjects’ confusions poorly: the best predictions were made, not by any of the spatial frequency similarities, but by similarities between the original letter stimuli input to the spatial frequency analysis programs. Previously reported feature analytic models seem to better describe the confusion patterns.

Early Comprehension of Mental Verbs: Think and Know

JOHNSON, CARL NILs, and MARATSOS, MICHAEL P. Early Comprehension of Mental Verbs: Think and Know. CHILD DEVELOPMENT, 1977, 48, 1743-1747. While it is a notable fact that preschool children use mental verbs which are remarkably abstract and complex in their implications, the developing understanding of such terms is not well understood. The present investigation was concerned with possible early confusion of internal mental states with external facts or events, such as confusing thinking with saying, and with the developing comprehension of different implications of the verbs think and know, namely, that knowing implies knowing a truth whereas thinking can equally be true or false. A story and question task was used in which characters were placed in contrasting states of knowledge and mistaken belief. Results revealed no consistent pattern of confusion of internal and external, and showed that 4-yearolds, but not 3-year-olds, comprehended the different implications of the terms.

Room reverberation effects on recognition of some consonant features

The effects of real room reverberation (T=0.75 sec) on speech sound recognition was investigated for young, normal-hearing subjects. Initial versus final and quiet versus reverberant results were significantly different. Place information was most affected, followed by frication. Stop and nasal information were next affected. Voicing was quite robust; and duration, sibilance, and liquid information were virtually unaffected. Several interesting confusion patterns emerged. Except for the relatively high susceptibility of frication, the effects of reverberation on speech recognition seem somewhat similar to that reported for masking and low-pass filtering. Reverberation appears to result in temporal smearing of the speech signal, so as to act as a speech-shaped masking noise. Also, short-duration events appear lengthened, and seem to be associated with stops being reported as fricatives; and relatively long temporal spreading of the fundamental and lower formants appears to be associated with stops being reported as nasals. Comparison of the results with previous data obtained with artificially induced reverberation revealed some qualitatively and quantitatively different effects.

Evaluating speech recognizers

Although automatic word recognition systems have existed for some twenty-five years there is still no suitable standard for evaluating their relative performances. Currently, the merits of two systems cannot be meaningfully compared unless they have been tested with at least the same vocabulary or, preferably, with the same acoustic samples. This paper develops a standard for comparing the performance of different recognizers on arbitrary vocabularies based on a human word recognition model. This standard allows recognition results to be normalized for comparison according to two intuitively meaningful figures of merit: 1) the noise level necessary to achieve comparable human performance and 2) the deviation of the pattern of confusions from human performance. Examples are given of recognizers evaluated in this way, and the role of these performance measures in automatic speech recognition and other related areas is discussed.

Adaptive Speech Audiometry and Speech Discrimination Space in Hearing-Impaired Subjects

Phonemic discrimination difficulties in a small group of patients with sensorineural hearing loss have been mapped using a new form of speech audiometry which is au automated, adaptive, audio-visual system, using a binary forced-choice response mode. Phonemic confusion patterns have been displayed in directed graph form and the cases presented here have been chosen to illustrate the potentialities of the test method for obtaining detailed information on the fine structure of speech reception in these patients.

Recognition of musical triad quality

Certain patterns of pitch, such as chords and scales, are central to the organization of traditional music. These patterns are identified by standardized nomenclature, and form a basis for analytical listening to music. The auditory processes involved in recognition may be revealed by comparing the ability of trained subjects to categorize these standard patterns. Perceptual confusion among standard, three-tone stimuli was studied in 19 musically trained subjects. A solid-state organ, interfaced to a digital computer, was used to present the stimuli; a terminal connected to the same computer was used to collect the responses. All subjects received the same stimuli, randomly accessed by the computer program. The subject indicated whether the signal presented was major, minor, augmented, or diminished. The three-note signals were presented in three manners: simultaneously, ascending (in pitch), and descending. All stimuli were presented at both O.2- and 0.1-sec note duration. Accuracy wan least for shorter duration and for simultaneous presentation. Different triad qualitites maintained essentially the same performance rank through the different modes of presentation. Patterns of confusion and response bias suggest the subjects' use of pre-existent auditory images.

Visual intelligibility of consonants: a lipreading screening test with implications for aural rehabilitation.

A lipreading screening test consisting of 100 consonant-vowel (CV) syllables was prepared on videotape and presented to subjected with normal hearing and vision. The results of visual consonant confusions revealed nine homophenous categories with high test-retest reliability. Based upon these categories, obtained under ideal viewing conditions, criterion levels of performance were specified in an attempt to determine the need for place of articulation instruction during lipreading training. Examples of confusion patterns from hearing-impaired observers are shown and implications for visual training are presented.

Initial and final phoneme confusions under reverberation

Initial and final consonant recognition confusions were investigated under reverberation times (RTs) of 0 to 3 sec. Mean recognition scores decreased with RT in both phoneme positions, but considerably more for final consonants. Initial phoneme confusion matrices were minimally affected by reverberation, with most errors being pure place confusions (/ptk/, /bdg/, /mn/, etc. ). Other initial confusions were randomly distributed, except for some nasals for voiced fricatives at 1‐sec RT. Final consonant confusions at 0‐sec RT also fell into place groupings, with other confusions randomly distributed. Under reverberation, however, while many final confusions were within place groups, several confusion patterns developed. Most frequently, these were substitutions of nasals for other phonemes. Voiceless fricatives were often perceived as voiceless stops at 1–3 sec RTs. At 3‐sec RT, voiced stops were often substituted in place of other phonemes. Alveolars (/tdn/) accounted for the greatest number of confusion responses. Resistance to reverberation distortion differed between initial and final consonants and between RTs. Initial phoneme groups were similar at 1‐sec RT, but the nasals and liquids were most susceptable to distortion at longer RTs. Among final phonemes, nasals and voiceless stops were the most robust, while fricatives were most susceptable to distortion. A basis for the confusion matrices under reverberation is suggested, and the data are compared with findings under noise, filtering, and other types of distortion. [Supported by VA.]

Acoustic confusion of digits in memory and recognition

It has been shown that, contrary to current opinion, digits do not confuse arbitrarily and that similar patterns of confusion are observed in memory and auditory recognition tasks. Two alternative measures of stimulus similarity are compared.

Consonant confusions in patients with sensorineural hearing loss.

Consonant confusion matrices were obtained from 22 outpatient listeners with sensorineural hearing loss for four sets of CV and VC nonsense syllables, presented monaurally at SRT + 40 DB. Testing was typically conducted for six hours on each of two separate days. Overall performance and patterns of confusions were stable over time. Analysis of the matrices in terms of phonological features indicated that the patterns of consonant confusions varied both with degree and configuration of the subject's loss. Scaling of intersubject similarity using a pairwise multidimensional scaling analysis resulted in consistent classification of subjects according to audiometric configuration into three groups-essentially normal hearing, flat or rising audiograms, and high-frequency hearing losses.