Confusable Pairs Research Articles

Voice biometrics security: Extrapolating false alarm rate via hierarchical Bayesian modeling of speaker verification scores

How secure automatic speaker verification (ASV) technology is? More concretely, given a specific target speaker, how likely is it to find another person who gets falsely accepted as that target? This question may be addressed empirically by studying naturally confusable pairs of speakers within a large enough corpus. To this end, one might expect to find at least some speaker pairs that are indistinguishable from each other in terms of ASV. To a certain extent, such aim is mirrored in the standardized ASV evaluation benchmarks, for instance, the series of speaker recognition evaluation (SRE) organized by the National Institute of Standards and Technology (NIST). Nonetheless, arguably the number of speakers in such evaluation benchmarks represents only a small fraction of all possible human voices, making it challenging to extrapolate performance beyond a given corpus. Furthermore, the impostors used in performance evaluation are usually selected randomly. A potentially more meaningful definition of an impostor — at least in the context of security-driven ASV applications — would be closest (most confusable) other speaker to a given target.We put forward a novel performance assessment framework to address both the inadequacy of the random-impostor evaluation model and the size limitation of evaluation corpora by addressing ASV security against closest impostors on arbitrarily large datasets. The framework allows one to make a prediction of the safety of given ASV technology, in its current state, for arbitrarily large speaker database size consisting of virtual (sampled) speakers. As a proof-of-concept, we analyze the performance of two state-of-the-art ASV systems, based on i-vector and x-vector speaker embeddings (as implemented in the popular Kaldi toolkit), on the recent VoxCeleb 1, and 2 corpora, containing a total of 7365 speakers. We fix the number of target speakers to 1000, and generate up to N=100,000 virtual impostors sampled from the generative model. The model-based false alarm rates are in a reasonable agreement with empirical false alarm rates and, as predicted, increase substantially (values up to 98%) with N=100,000 impostors. Neither the i-vector or x-vector system is immune to increased false alarm rate at increased impostor database size, as predicted by the model.

A Seed-Based Method for Generating Chinese Confusion Sets

In natural language, people often misuse a word (called a “confused word”) in place of other words (called “confusing words”). In misspelling corrections, many approaches to finding and correcting misspelling errors are based on a simple notion called a “confusion set.” The confusion set of a confused word consists of confusing words. In this article, we propose a new method of building Chinese character confusion sets. Our method is composed of two major phases. In the first phase, we build a list of seed confusion sets for each Chinese character, which is based on measuring similarity in character pinyin or similarity in character shape. In this phase, all confusion sets are constructed manually, and the confusion sets are organized into a graph, called a “seed confusion graph” (SCG), in which vertices denote characters and edges are pairs of characters in the form (confused character, confusing character). In the second phase, we extend the SCG by acquiring more pairs of (confused character, confusing character) from a large Chinese corpus. For this, we use several word patterns (or patterns) to generate new confusion pairs and then verify the pairs before adding them into a SCG. Comprehensive experiments show that our method of extending confusion sets is effective. Also, we shall use the confusion sets in Chinese misspelling corrections to show the utility of our method.

Tallman lettering as a strategy for differentiation in look-alike, sound-alike drug names: The role of familiarity in differentiating drug doppelgangers

Tallman lettering, capitalizing the dissimilar portions of easily confused drug names, is one strategy for reducing medication errors. We assessed the efficacy of Tallman lettering in a visually complex environment using a change detection method with healthcare providers and laypeople. In addition, the effect of familiarity with the drug name was assessed using a subset of responses collected from healthcare providers.Both healthcare providers and laypeople detected changes in confusable pairs of drug names more often (P < 0.0001) and more quickly (P < 0.05) when changes were presented in Tallman lettering, though the benefits were more pronounced for healthcare providers (p < 0.05). Familiarity with both drug names in a confusable pair mitigated the benefit of Tallman lettering. Results are discussed in terms of bottom-up and top-down attentional systems for processing of information in the context of the varied healthcare environments.

Brain correlates of experience-dependent changes in stimulus discrimination based on the amount and schedule of exposure.

One product of simple exposure to similar visual stimuli is that they become easier to distinguish. The early visual cortex and other brain areas (such as the prefrontal cortex) have been implicated in such perceptual learning effects, but the anatomical specificity within visual cortex and the relationship between sensory cortex and other brain areas has yet to be examined. Moreover, while variations in the schedule (rather than merely the amount) of exposure influence experience-dependent improvement in discrimination, the neural sequelae of exposure schedule have not been fully investigated. In an event-related fMRI study, participants were exposed to confusable pairs of faces, scenes and dot patterns, using either intermixed or blocked presentation schedules. Participants then performed same/different judgements with exposed and novel pairs of stimuli. Stimulus independent activation, which was correlated with experience-dependent improvement in discrimination, was seen in frontal areas (e.g. frontal and supplementary eye fields and dorsolateral prefrontal cortex) and in early visual cortex (V1-4). In all regions, the difference in activation between exposed and novel stimuli decreased as a function of the degree of discrimination improvement. Overall levels of BOLD activation differed across regions, consistent with the possibility that, as a consequence of experience, processing shifts from initial engagement of early visual regions to higher order visual areas. Similar relationships were observed when contrasting intermixed with blocked exposure, suggesting that the schedule of exposure primarily influences the degree of, rather than the mechanisms for, discrimination performance.

Do medial temporal lobe regions play a domain-specific or domain-independent role in perceptual learning performance?

Event Abstract Back to Event Do medial temporal lobe regions play a domain-specific or domain-independent role in perceptual learning performance? Matthew E. Mundy1*, Paul Downing2, Rob Honey3, Dominic Dwyer3 and Kim Graham3 1 Monash University, Australia 2 Bangor University, United Kingdom 3 Cardiff University, United Kingdom Background It is contentious whether structures in the human medial temporal lobe (MTL) play a domain-specific or domain-independent role in learning and memory. Perceptual learning has been used as a tool to investigate this critical cognitive junction. Methods In a series of event related fMRI experiments, participants made repeated same/different judgements to previously seen and initially novel confusable pairs of dot patterns, faces and complex scenes. Using a series of orthogonal, and independent, functional localisers, clusters of stimulus-selective, novelty-sensitive voxels were identified in two medial temporal regions (perirhinal cortex and posterior hippocampus), and two extrastriate regions (fusiform face area, FFA, and parahippocampal place area, PPA). We asked how activity in these regions was influenced by discrimination accuracy and by trial repetition (e.g., adaptation). Results In contrast to FFA and PPA, which only cared about preferred category, activity in perirhinal cortex and posterior hippocampus predicted discrimination accuracy for faces and scenes, respectively. MTL regions also adapted less rapidly than extrastriate areas over trial repetition, with difference emerging between extrastriate and MTL regions after 8 repetitions. Structural differences were also recorded in these regions, along with caudate and thalamic areas; the size of these differences correlated with performance. Discussion These findings, supported by our recent work with MTL lesioned patients and a second fMRI investigation directly modulating stimulus ambiguity, show that domain-specific patterns of responding in the human brain are not just restricted to extrastriate cortex, and highlight a key role for perirhinal cortex and posterior hippocampus, but not FFA and PPA, in storing feature ambiguous representations of faces and scenes, respectively. Keywords: faces, scenes, Perceptual Learning, Memory, Amnesia, Medial Temporal Lobes, fMRI Conference: ACNS-2013 Australasian Cognitive Neuroscience Society Conference, Clayton, Melbourne, Australia, 28 Nov - 1 Dec, 2013. Presentation Type: Oral Topic: Memory Citation: Mundy ME, Downing P, Honey R, Dwyer D and Graham K (2013). Do medial temporal lobe regions play a domain-specific or domain-independent role in perceptual learning performance?. Conference Abstract: ACNS-2013 Australasian Cognitive Neuroscience Society Conference. doi: 10.3389/conf.fnhum.2013.212.00192 Copyright: The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers. They are made available through the Frontiers publishing platform as a service to conference organizers and presenters. The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated. Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed. For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions. Received: 15 Oct 2013; Published Online: 25 Nov 2013. * Correspondence: Dr. Matthew E Mundy, Monash University, Melbourne, Australia, matthew.mundy@monash.edu Login Required This action requires you to be registered with Frontiers and logged in. To register or login click here. Abstract Info Abstract The Authors in Frontiers Matthew E Mundy Paul Downing Rob Honey Dominic Dwyer Kim Graham Google Matthew E Mundy Paul Downing Rob Honey Dominic Dwyer Kim Graham Google Scholar Matthew E Mundy Paul Downing Rob Honey Dominic Dwyer Kim Graham PubMed Matthew E Mundy Paul Downing Rob Honey Dominic Dwyer Kim Graham Related Article in Frontiers Google Scholar PubMed Abstract Close Back to top Javascript is disabled. Please enable Javascript in your browser settings in order to see all the content on this page.

Error analysis to improve the speech recognition accuracy on Telugu language

Speech is one of the most important communication channels among the people. Speech Recognition occupies a prominent place in communication between the humans and machine. Several factors affect the accuracy of the speech recognition system. Much effort was involved to increase the accuracy of the speech recognition system, still erroneous output is generating in current speech recognition systems. Telugu language is one of the most widely spoken south Indian languages. In the proposed Telugu speech recognition system, errors obtained from decoder are analysed to improve the performance of the speech recognition system. Static pronunciation dictionary plays a key role in the speech recognition accuracy. Modification should be performed in the dictionary, which is used in the decoder of the speech recognition system. This modification reduces the number of the confusion pairs which improves the performance of the speech recognition system. Language model scores are also varied with this modification. Hit rate is considerably increased during this modification and false alarms have been changing during the modification of the pronunciation dictionary. Variations are observed in different error measures such as F-measures, error-rate and Word Error Rate (WER) by application of the proposed method.

Lexical neighborhoods and phonological confusability in cross-dialect word recognition in noise

AbstractLexical neighborhood density is a well-known factor affecting phonological categorization in spoken word recognition. The current study examined the interaction between lexical neighborhood density and dialect variation in spoken word recognition in noise. The stimulus materials were real English words produced in two regional American English dialects. To manipulate lexical neighborhood density, target words were selected so that predicted phonological confusions across dialects resulted in real English words in the word-competitor condition and did not result in real English words in the nonword-competitor condition. Word and vowel recognition performance were more accurate in the nonword-competitor condition than the word-competitor condition for both talker dialects. An examination of the responses to specific vowels revealed the role of dialect variation in eliciting this effect. When the predicted phonological confusions were real lexical neighbors, listeners could respond with either the target word or the confusable minimal pair, and were more likely than expected to produce a minimal pair differing from the target by one vowel. When the predicted phonological confusions were not real words, however, the listeners exhibited less lexical competition and responded with the target word or a minimal pair differing by one consonant.

Which words are hard to recognize? Prosodic, lexical, and disfluency factors that increase speech recognition error rates

Despite years of speech recognition research, little is known about which words tend to be misrecognized and why. Previous work has shown that errors increase for infrequent words, short words, and very loud or fast speech, but many other presumed causes of error (e.g., nearby disfluencies, turn-initial words, phonetic neighborhood density) have never been carefully tested. The reasons for the huge differences found in error rates between speakers also remain largely mysterious. Using a mixed-effects regression model, we investigate these and other factors by analyzing the errors of two state-of-the-art recognizers on conversational speech. Words with higher error rates include those with extreme prosodic characteristics, those occurring turn-initially or as discourse markers, and doubly confusable pairs: acoustically similar words that also have similar language model probabilities. Words preceding disfluent interruption points (first repetition tokens and words before fragments) also have higher error rates. Finally, even after accounting for other factors, speaker differences cause enormous variance in error rates, suggesting that speaker error rate variance is not fully explained by differences in word choice, fluency, or prosodic characteristics. We also propose that doubly confusable pairs, rather than high neighborhood density, may better explain phonetic neighborhood errors in human speech processing.

Evolutionary models of color categorization I Population categorization systems based on normal and dichromat observers

The evolution of color categorization is investigated using artificial agent population categorization games, by modeling observer types using Farnsworth-Munsell 100 Hue Test performance to capture human processing constraints on color categorization. Homogeneous populations of both normal and dichromat agents are separately examined. Both types of populations produce near-optimal categorization solutions. While normal observers produce categorization solutions that show rotational invariance, dichromats' solutions show symmetry-breaking features. In particular, it is found that dichromats' local confusion regions tend to repel color category boundaries and that global confusion pairs attract category boundaries. The trade-off between these two mechanisms gives rise to population categorization solutions where color boundaries are anchored to a subset of locations in the stimulus space. A companion paper extends these studies to more realistic, heterogeneous agent populations [J. Opt. Soc. Am. A26, 1424-1436 (2009)].

Relative legibility and confusions of letter acuity targets in the peripheral and central retina.

Patients with macular disease may image optotypes at a peripheral retinal locus during visual acuity testing. In this study, we asked if the relative legibility and confusions between letters are similar in the fovea and at 10 degrees in the peripheral retina. Twenty five upper-case alphabet letters (all except "I"), constructed to the same specifications as the Sloan letters, were presented one at a time on a computer monitor at the fovea and at 10 degrees in the temporal, superior and superior-temporal retina. The range of relative legibility at peripheral loci was generally larger than in the fovea, for all letters and for the subset of 10 Sloan letters. Although many confusion pairs were similar in the fovea and periphery, additional confusion pairs, preferentially involving curved letters, occurred uniquely in the periphery. The increased range of relative legibility for letter targets in the peripheral retina underscores the importance of using letter-by-letter scoring to obtain precise measures of visual acuity in patients without central vision.

Lexical Retrieval and Selection Processes: Effects of Transposed-Letter Confusability

Three experiments investigated performance for words which differ from another word only by the transposition of two letters (e.g.,salt, slat). In Experiment 1, high frequency words from transposed-letter (TL) confusable pairs were responded to more slowly than carefully matched control words in both the lexical decision and word naming task. Low frequency TL words were responded to less accurately than control words in the naming but not the lexical decision task. Experiment 2 replicated the naming data of Experiment 1 and also revealed that naming accuracy for TL word targets was reduced when they were preceded by a brief masked presentation of their confusable mate. Experiment 3 provided a third replication of the impaired naming performance for TL target words and demonstrated that the effect was insensitive to concurrent dual task demands. These TL confusability effects provide strong constraints that can contribute to evaluation and specification of current models of visual word recognition.

The effect of adaptation on the differential sensitivity of the S-cone color system

This paper presents a psychophysical dissection of the S-cone color system. Experiments were guided by a skeletal model that assumed a first stage consisting of S-, M- and L-cones, and a second stage of the opponent combination of the 5 and L + M signals. The response of the S-cone system was isolated by measuring difference thresholds between lights that were equiluminant tritanopic confusion pairs and thus differed only in S-cone excitation. Two types of mechanisms that control sensitivity in the S-cone system were identified: (i) static mechanisms that have a restricted range and thus limit discrimination to a small range of inputs; and (ii) adaptive mechanisms that change the state of the system in response to changes in steady illumination, so that the system is sensitive to small changes from the adapting light. These mechanisms were localized by lights that stimulated the S-cone system while keeping the signal constant at either the S, the L + M, or the post-opponent stage. The response function of the static mechanism was estimated by measuring difference thresholds at judgment points other than the steady adapting light. This procedure was repeated at a number of adaptation lights to examine the properties of adaptive mechanisms. The data were consistent with an elaborated model that included identical multiplicative gain control mechanisms in the S and L + M pre-opponent branches, and a post-opponent static sigmoidal nonlinearity with different amounts of compression for positive and negative opponent inputs.

A Perceptual Study of Mandarin Tones 2 and 3

Due to their similar concave shapes, Tones 2 and 3 are said to be the most confusable pair in Mandarin, and no reliable perceptual cues have been found. The purpose of this perceptual study was to find (1) the perceptual cues for discriminating these two tones, and (2) the possible causes of listeners' misidentifications. Experiment 1 was designed to test the hypothesis that the timing of the turning point may constitute a perceptual cue. Two fundamental frequency continua were superimposed on the syllable [wu] and the location of the turning point was manipulated in each. A binary forced-choice identification test was given to Mandarin listeners. The results show that the distinction between Tones 2 and 3 is cued by the timing of the turning point, which is correlated with the degree of the initial fall. In Experiment 2, ten minimal pairs of Tones 2 and 3 spoken by four native speakers were presented to Mandarin listeners for labeling. Acoustic analysis of the mis-labeled tokens indicates that mistakes occurred when the correlation between the timing of the turning point and the degree of the initial fall was violated.

Human tactile pattern recognition: active versus passive touch, velocity effects, and patterns of confusion

1. Subjects without any previous experience in a tactile psychophysics task participated in a study of tactile letter recognition employing active and passive touch. In the active task, subjects reached through a curtain and examined embossed letters with horizontal, unidirectional finger strokes. In the passive task, subjects sat with their arms and hands immobilized while a rotating drum stimulator pressed the embossed letters onto the right index finger. The stimulus conditions in the passive task were identical to those used in neurophysiological experiments with monkeys. 2. A survey of 40 naive subjects who were not screened in any way showed a wide range of performance levels. There was no difference between the subjects in the active and passive tasks, either in overall mean percent correct scores, which were 49.0 and 50.7%, respectively or in the percent correct scores for individual letters whose product-moment correlation coefficient was 0.94. The active and passive groups, which contained 25 and 15 members, respectively, had no members in common. 3. Videotapes of the finger movements of eight subjects in the active task showed a characteristic V-shaped velocity profile (velocity vs. lateral position) starting at approximately 100 mm/s at the left-hand edge of the plate containing the embossed letter, decelerating to a minimum when the center of the finger was directly over the letter, and then accelerating away from the letter. The average minimum scanning velocity was 17 mm/s. 4. Scanning velocity had no significant effect on performance in the passive task between 20 and 40 mm/s. An increase to 80 mm/s produced a 16% decline in percent correct identifications. 5. Learning effects were evident across sessions even though subjects were given no feedback or training. The increase in mean percent correct judgments averaged 4% per session, which lasted for approximately 1 h. 6. Data from 64 subjects were pooled for detailed comparison of identification patterns in active and passive touch. The results were analyzed and found to be consistent with the hypothesis that the identification and confusion probabilities are identical in the two modes. We conclude that there is no difference between active and passive touch in form recognition when the stimulus pattern is smaller than a finger pad. 7. Data from all experiments were pooled to produce a single confusion matrix with 324 presentations per letter. The majority of erroneous responses are grouped in a small number of confusion pairs and the majority of those confusion pairs are strongly asymmetric. The probable neural mechanisms of some confusion patterns are discussed.

Positional differences in performance on members of confusable and nonconfusable letter pairs.

Positional differences in performance on members of confusable and nonconfusable letter pairs.

Positional differences in performance on members of confusable and nonconfusable letter pairs.

A parafoveally presented letter is more accurately identified when flanked by a letter to its foveal side than when flanked by one to its peripheral side, but only if the two letters are nonconfusable. With confusable letters there is no such relative position effect. Four experiments indicated that the basis for this confusability-asymmetry interaction is not criterion or response bias, but rather that it occurs earlier in visual processing. In Experiment 1 the interaction was found when only one pair member was reported, thus eliminating response bias requiring the report of both letters as the source of the effect. In Experiment 2 the data were subjected to signal detection analysis, and the interaction persisted. In Experiment 3 pair members were presented simultaneously or in rapid sequence, and the interaction was found only with simultaneous presentations. In Experiment 4 letters were used with upper- and lowercase counterparts that were quite different in shape. Uppercase letters that were most and least confusable for each subject were paired for presentation in their upper case form or in mixed-case form. The interaction occurred only with upper-case pairs.

The role of spatial redundancy in grapheme recognition: perception or inference?

The question under investigation is whether spatial redundancy operates by restricting the number of valid alternatives at each serial position or whether it serves a perceptual function by keeping visually confusable graphemes from appearing in the same array and/or maximizing the distance between such graphemes when they do occur in the same array. The first experiment used the same-different task to establish a confusion matrix for 12 non-linguistic symbols. The spatial relation between the members of two pairs of confusable symbols was manipulated in a second experiment, which had subjects look for the presence or absence of a predetermined target symbol in single six-symbol linear arrays. Facilitation was found to be dependent upon the perceptual consequences of serial position constraint rather than upon the constraint per se. The constraint was most effective when it served to prevent members of visually confusable pairs from occurring in the same array. Since spatial redundancy may operate at the feature extraction stage by minimizing competition for the same sets of feature detectors and since words are likely to be high in spatial redundancy, the possibility is raised that part of the word superiority effect may be due to visual factors rather than to linguistic context.

Note on Validity of a Mechanical Measure of Inter-Letter Similarity

Two studies recently published in this Journal (Dunn-Rankin, et al., 1968; Kuennapas & Janson, 1969) have presented alternative measures of inter-letter similarity (ILS) between lower-case alphabetic letters. The two measures-which we shall call the mechanical measure (MM) and the subject-based measure (SBM)--differed in several respects. Dunn-Rankin, et al.'s MM, for example, used letters of the English alphabet while Kuenn a p s and Janson's SBM used the Swedish alphabet, which omits the letter 'w' and includes the letters 'i'. ' a , and '6'. In addition, MM was applied to a particular type of lower-case print (Century Schoolbook) whereas SBM employed lower-case letters of 'simple shape' designed by Es themselves. Most importantly, however, SBM used a sample of 57 Ss (presumably Swedish) to rate the perceived inter-letter similarity on a 0 to 100 scale; by contrast, MM based its ILS measure on the ratio of common surface area shared by two letters to the combined remaining areas. In both studies Es factor analyzed their matrices of ILS scores in an attempt to isolate critical features which determine the degree of resemblance between letters. In general, the resulting patterns of factor loadings are subjectively quite comparable. Some applications, however, might require the use of ILS measures in their own right. In such cases, it becomes important to know whether a technique such as MM can validly represent ILS as measured by a more time-consuming procedure such as SBM. Dunn-Rankin attempted to validate MM by applying it to English capital letters and comparing the results with a subject-based confusion matrix developed by Gibson, et a/. (1963) using nursery-school children. Among the 36 highest-ranhng confusion pairs, Dunn-Rankin found a rank-order correlation of .50 with his own MM scores (DunnRankin, er al., 1968, p. 660). The ILS matrix of Kuennapas and Janson, however, permits a more direct test of MM's validity. A correlation analysis of the 300 lower-case letter pairs shared by the English and Swedish alphabets yields an r of .77 between MM and SBM. (Visual inspection of the scatter plot did not suggest any departure from linearity in this relationship.) Comparable correlacioas, performed separately on MM's 37 highest-ranked and SBM's 36 highest-ranked pairs, are .67 and .49 respectively. These three criteria of MM's validity compare favorably with the .50 rank correlation originally claimed by Dunn-Rankin, et al. In our view, the previously noted differences in alphabet nationality and type face may easily account for much of the remaining variance between MM and SBM. Pending further evidence, we may therefore attribute considerable validiry to Dunn-Ranlun, et al.'s mechanical measure of inter-letter similarity.

Computer Program to Aid in the Selection and Evaluation of Vocabularies for Word Recognizers

The performance of a word recognizer depends upon the vocabulary that it attempts to recognize. A computer program has been written to aid in the selection and evaluation of vocabularies for word recognizers. The program operates by predicting the performance of a given recognizer on a given vocabulary. The words in the vocabulary are treated as strings of phonemes, and it is assumed that the word recognizer being evaluated operates in terms of such strings. The program searches the vocabulary for word pairs that would be confused with each other if the recognizer were unable to make certain discriminations. When a confusable pair is found, it may be added to a list of confusions for later printing, or one word of the pair may be deleted from the vocabulary. The latter mode of operation results in a reduced vocabulary on which the recognizer should perform well. Results are presented for a vocabulary of about 1000 common words.