Nonsense Utterances Research Articles

Preschool[-aged] children's use of perceptual features to identify spoken languages

One major question in language development is when and how young language learners identify what language(s) they are hearing. Infants can discriminate languages when at least one is familiar. We have previously demonstrated that regardless of language background, 3- to 5-year-olds accurately associate two languages with two individuals (cartoons), but how they accomplish this is unclear. Here, we test whether bilingual (n = 32) and monolingual (n = 32) children associate perceptual features versus comprehensibility with individuals. During learning, children are familiarized with two cartoon characters: one speaks Spanish sentences; the other, English. At test, children see both characters and hear four trial types: English sentences; Spanish sentences; English-like nonsense; Spanish-like nonsense. Sense and nonsense sentences are acoustically matched via bidirectional cross-splicing. On each test trial, the child points to the character who they thought spoke, while eye-movements are being recorded. If children associate language-specific phonetic cues with talkers, they will choose English-speaking characters for English-like nonsense, and Spanish-speaking characters for Spanish-like nonsense. However, English monolinguals may associate characters with (non)comprehensibility. If so, they will choose the Spanish-speaking character for all nonsense utterances. Findings will reveal how much children use low-level perceptual features in language recognition, and whether this differs between bilinguals and monolinguals.One major question in language development is when and how young language learners identify what language(s) they are hearing. Infants can discriminate languages when at least one is familiar. We have previously demonstrated that regardless of language background, 3- to 5-year-olds accurately associate two languages with two individuals (cartoons), but how they accomplish this is unclear. Here, we test whether bilingual (n = 32) and monolingual (n = 32) children associate perceptual features versus comprehensibility with individuals. During learning, children are familiarized with two cartoon characters: one speaks Spanish sentences; the other, English. At test, children see both characters and hear four trial types: English sentences; Spanish sentences; English-like nonsense; Spanish-like nonsense. Sense and nonsense sentences are acoustically matched via bidirectional cross-splicing. On each tes...

Utterance-final position and pitch marking aid word learning in school-age children.

We investigated the effects of word order and prosody on word learning in school-age children. Third graders viewed photographs belonging to one of three semantic categories while hearing four-word nonsense utterances containing a target word. In the control condition, all words had the same pitch and, across trials, the position of the target word was varied systematically within each utterance. The only cue to word–meaning mapping was the co-occurrence of target words and referents. This cue was present in all conditions. In the Utterance-final condition, the target word always occurred in utterance-final position, and at the same fundamental frequency as all the other words of the utterance. In the Pitch peak condition, the position of the target word was varied systematically within each utterance across trials, and produced with pitch contrasts typical of infant-directed speech (IDS). In the Pitch peak + Utterance-final condition, the target word always occurred in utterance-final position, and was marked with a pitch contrast typical of IDS. Word learning occurred in all conditions except the control condition. Moreover, learning performance was significantly higher than that observed with simple co-occurrence (control condition) only for the Pitch peak + Utterance-final condition. We conclude that, for school-age children, the combination of words' utterance-final alignment and pitch enhancement boosts word learning.

Domain generalisation in artificial language learning

Many languages have restrictions on word-final segments, such as a requirement that any word-final obstruent be voiceless. There is a phonetic basis for such restrictions at the ends of utterances, but not the ends of words. Historical linguists have long noted this mismatch, and have attributed it to an analogical generalisation of such restrictions from utterance-final to word-final position. To test whether language learners actually generalise in this way, two artificial language learning experiments were conducted. Participants heard nonsense utterances in which there was a restriction on utterance-final obstruents, but in which no information was available about word-final utterance-medial obstruents. They were then tested on utterances that included obstruents in both positions. They learned the pattern and generalised it to word-final utterance-medial position, confirming that learners are biased toward word-based distributional patterns.

Comparison across languages using the multilingual Matrix Test: Which language is best to survive in a cocktail party?

The Matrix test (i.e., sentence test with fixed syntactic structure, but ten alternative words in each position which may lead to nonsense utterances) has the potential to overcome the inherently language-dependent incompatibilities of speech audiology. It is meanwhile available (with varying degree of supportive data) in Swedish, German, Danish, Dutch, American English, British English, French, Polish, Turkish, Spanish, Italian, Persian, Arabian, Finnish, and Russian. Several measures have been taken to make the tests as efficient, reliable and comparable across different languages as possible and to establish a de-facto standard. Using the Matrix concept it is also possible to estimate the “communication efficiency” of the different languages for this kind of sentences against each other. To eliminate the influence of the individual speaker, recordings with accent-free bilingual speakers (German-Russian and German-Spanish) were used to assess the respective speech reception threshold (SRT) for native listeners using stationary and fluctuating background noise. The results show both an inter-speaker and inter-language effect in the order of 3 dB. The latter is larger between German and Spanish than between German and Russian. The origin of these effects (such as long-term speech spectrum and the relative information content of consonants and vowels) will be discussed.

Improving automatic speech recognition by learning from human errors

This work presents a series of experiments that compare the performance of human speech recognition (HSR) and automatic speech recognition (ASR). The goal of this line of research is to learn from the differences between HSR and ASR and to use this knowledge to incorporate new signal processing strategies from the human auditory system in automatic classifiers. A database with noisy nonsense utterances is used both for HSR and ASR experiments with focus on the influence of intrinsic variation (arising from changes in speaking rate, effort, and style). A standard ASR system is found to reach human performance level only when the signal-to-noise ratio is increased by 15 dB, which can be seen as the human–machine gap for speech recognition on a sub-lexical level. The sources of intrinsic variation are found to severely degrade phoneme recognition scores both in HSR and in ASR. A comparison of utterances produced at different speaking rates indicates that temporal cues are not optimally exploited in ASR, which results in a strong increase of vowel confusions. Alternative feature extraction methods that take into account temporal and spectro-temporal modulations of speech signals are discussed.

Different motor strategies for increasing speaking rate: Data and modeling

Different motor strategies for increasing speaking rate: data and modeling EMG, kinematic and acoustic signals were recorded from two male subjects as they pronounced multiple repetitions of simple nonsense utterances. The resulting data indicate that the two subjects employed different motor strategies to increase speaking rate. When speaking faster, S1 significantly increased the size of the articulatory target region for his tongue movements, increased the speed of the tongue movements and the rate of EMG rise somewhat, while decreasing the movement duration significantly and movement distance slightly. In contrast, at the fast rate, S2 had the same size articulatory target region and rate of EMG rise as at the normal rate, but decreased the speed, distance, and duration of tongue movement slightly. Each subject had similar dispersions of acoustic targets in F1–F2 space at fast versus normal rates, but both shifted target centroids toward the center of the vowel space at the fast rate. Simulations with a biomechanical model of the vocal tract show how modulations of motor commands may account for such effects of speaking rate on EMG, kinematics, and acoustic outputs. [Work supported by NIDCD, NIH.]

Working memory and sentence comprehension of Hong Kong Chinese children with specific language impairment

Children with Specific Language Impairment present with delayed language development, but do not have a history of hearing impairment, mental deficiency, or associated social or behavioral problems. Non-word repetition was suggested as an index to reflect the capacity of phonological working memory. There is a paucity of such studies among Hong Kong Chinese children. This preliminary study aimed to examine the relationship between phonological working memory and Specific Language Impairment, through the processes of non-word repetition and sentence comprehension, of children with Specific Language Impairment and pre-school children with normal language development. Both groups of children were screened by a standardized language test. A list of Cantonese (the commonest dialect used in Hong Kong) multisyllabic nonsense utterances and a set of 18 sentences were developed for this study. t-tests and Pearson correlation were used to study the relationship between non-word repetition, working memory and specific language impairment. Twenty-three pre-school children with Specific Language Impairment (mean age = 68.30 months; SD = 6.90) and another 23 pre-school children (mean age = 67.30 months; SD = 6.16) participated in the study. Significant difference performance was found between the Specific Language Impairment group and normal language group in the multisyllabic nonsense utterances repetition task and the sentence comprehension task. Length effect was noted in Specific Language Impairment group children, which is consistent with the findings of other literature. In addition, correlations were also observed between the number of nonsense utterances repeated and the number of elements comprehended. Cantonese multisyllabic nonsense utterances might be worth further developing as a screening tool for the early detection of children with Specific Language Impairment.

Free choice task effects in cross-linguistic stop perception

This paper examines the identification of stop place and secondary articulation using a free choice task. Russian syllable-initial and syllable-final stops /p pj t tj/ in nonsense utterances were presented to Russian and Japanese listeners (N=30). Correct identification rates for place and secondary articulation of the target consonants were determined based on written responses (in Cyrillic or Katakana). Both groups of listeners showed better identification of syllable-initial stops compared to syllable-final stops. Among the consonants, /p/ was identified better, and /pj/ was identified worse than the other stops. Native listeners performed better than non-native listeners. The overall correct identification rates were lower than (yet strongly correlated with) the rates previously obtained with the same stimuli using a forced choice phoneme identification task. The lower identification rates in the current study can be explained in part by the errors involving the segmentation and syllabification of palatalized stops. Thus, the palatal articulation of the syllable-final palatalized /pj/ was often interpreted as independent of the stop (e.g., /tapj api/ rendered as /taj papi/ or /tjap api/). It is concluded that the free choice task can successfully complement the forced choice task, providing additional information about the perception of secondary palatalization. [Supported by SSHRC.]

Coarticulation patterns in children with developmental apraxia of speech

The aim of this study was to enhance our insight into the underlying deficit in developmental apraxia of speech (DAS). In particular, the involvement of planning and/or programming of speech movements in context was tested by analysing coarticulatory cohesion. For this purpose, second formant frequency measurements were conducted in repetitions of nonsense utterances ([ CV]C=/s,x,b,d/; V=/i,a,u/), and compared across nine children with DAS, six normally speaking (NS) children and six adult women. The results showed both intra- and intersyllabic anticipatory coarticulation in NS children and adult women, in which the intersyllabic coarticulation was stronger in NS children than in adult women. The children with DAS showed more variability as compared to NS children, made, on average, less distinction between the vowels, and showed individually idiosyncratic coarticulation patterns. These results are discussed in the light of a delay as well as a deviance of speech development in children with DAS.

Speech production: Thirty years after

Measurements have been made on a corpus of bisyllabic nonsense utterances spoken by three adult male talkers on two occasions separated by 30 years. The 1960 and 1990 recordings were processed similarly so that comparisons could be made directly. Some primary and secondary characteristics of vowels were measured: duration, fundamental frequency and the frequencies of F1 and F2. The durations of prestressed consonants were measured, and vowel–consonant amplitude ratios computed. Long‐term spectra of the talkers were examined and some informal listening comparisons were made. Spectrograms of some samples were studied, and statistical analyses, when appropriate, were made. Results indicate that the talkers’ productions were remarkably consistent over 30 years. Over time the talkers’ mean vocalic duration (10 vowels) shortened (6 ms); their mean vocalic fundamental frequency rose (7 Hz); their mean frequencies of F1 and F2 (10 vowels) descended (15 and 26 Hz, respectively); their mean (initial) fricative duration lengthened (12 ms); and, their mean (initial) stop duration lengthened (hold portion, 15 ms; release portion, 5 ms). Interactions were usually significant: individual talkers’ changes in duration, formant frequencies; and fundamental frequency were not identical. Patterns of duration and fundamental frequency in the set of vowels, however, were stable over time.

Lower pharyngeal wall coarticulation in VCV syllables.

Speech movements of the lower pharyngeal wall were recorded in two subjects using pulsed-echo ultrasound. The focus of the study was the pattern of coarticulation of pharyngeal wall movements. Using nonsense utterances as test material, both anticipatory and carryover coarticulatory effects were observed. The identity of the final vowel in VCV sequences affected the kinematic characteristics of the initial VC transition. Both the amplitude and the duration of the movement between the initial vowel and the consonant were greater when the final vowel was /u/ rather than /a/. Similarly, the initial vowel affected the kinematic characteristics of the final CV transition. The amplitude of the movement from the consonant to the final vowel was greater with the initial vowel /u/ as opposed to /a/. The coarticulatory patterns observed in this study are similar to those previously reported for the tongue dorsum and upper pharynx [Parush et al., J. Acoust. Soc. Am. 74, 1115-1125 (1983); Parush and Ostry, J. Acoust. Soc. Am. 80, 749-756 (1986)].

Magnitude-Estimation Scaling of Speech Intelligibility: Effects of Listeners' Experience and Semantic-Syntactic Context

Fifteen speech-language pathologists with extensive experience judging speakers' intelligibility and 15 control subjects with no such previous experience provided magnitude-estimation responses for two sets of nine audiotaped speech samples. These samples were three utterances composed of a group of 17 words that contained all the consonant phonemes of English. These words were arranged to form a set of either meaningful or nonsense utterances. Nine separate versions of both the meaningful and nonsense utterances were created by systematically increasing the number of phonemes produced incorrectly on each of the nine recordings. The analysis indicated no significant difference between the magnitude-estimation scaling responses of experienced and inexperienced listeners. A significant over-all difference was found for listeners' responses to meaningful versus nonsense utterances. The advantages of magnitude-estimation scaling as a measure of speakers' intelligibility are discussed.

Magnitude estimation scaling of intelligibility by experienced and inexperienced listeners

Fifteen experienced listeners and 15 inexperienced listeners provided magnitude estimation scaling responses to indicate the intelligibility of two sets of nine audiotaped speech samples. These samples consisted of three utterances made up of 17 words and contained all of the consonant phonemes of English. These words were arranged to form a set of either meaningful or nonsense utterances. Nine separate versions of both the meaningful and nonsense utterances were created by systematically increasing the number of phonemes produced incorrectly on each of the nine recordings. Results indicated no significant difference between the magnitude estimation scaling responses of experienced and inexperienced listeners. Explanations for the results of this study and the advantages of magnitude estimation scaling as a measure of speaker intelligibility are discussed.

On enhancement of spectral contrast in speech for hearing-impaired listeners

A digital processing method is described for altering spectral contrast (the difference in amplitude between spectral peaks and valleys) in natural utterances. Speech processed with programs implementing the contrast alteration procedure was presented to listeners with moderate to severe sensorineural hearing loss. The task was a three alternative (/b/,/d/, or /g/) stop consonant identification task for consonants at a fixed location in short nonsense utterances. Overall, tokens with enhanced contrast showed moderate gains in percentage correct stop consonant identification when compared to unaltered tokens. Conversely, reducing spectral contrast generally reduced percent correct stop consonant identification. Contrast alteration effects were inconsistent for utterances containing /d/. The observed contrast effects also interacted with token intelligibility.

EMG analysis of speech production of patients with Friedreich disease

EMG activity from the orbicularis oris (superior and inferior), mentalis, depressor labii inferior (four lip muscles) and anterior belly of the digastric (one jaw muscle) was recorded during nonsense utterances produced at two speaking rates, conversational and fast, by 13 patients with Friedreich disease. The data were compared with those obtained from two normal subjects. At a conversational speaking rate, all patients showed a prolongation of anticipatory muscular activities and overall muscular activities. This abnormality might be an elemental feature of dysmetria. Hypotonia could also be considered as an explanation. At a conversational speaking rate, muscular synergia was better preserved than at a fast speaking rate. However, only five of 13 patients showed a normal EMG pattern with reciprocity between the muscular activities for rhythmical opening/closing movements of lips and jaw (/papapa/). For the same utterance produced at a fast speaking rate, no patient showed a normal EMG pattern. This abnormality might reflect dysdiadochokinesia. Production of the syllable /ba/ in response to an auditory signal showed a delay in the initiation of muscular activities which might be explained by hypotonia.

Recognition of intervocalic stops in continuous speech using context-dependent HMMs

In this work the design and evaluation of the recognition performance of context dependentHidden Markov Models (HMMs) for the intervocalic voiced and unvoiced stops is described. The phoneme HMMs are context-dependent in order to account for coarticulatory effects. Continuous probability density functions are used for the out putvectors. Initial model parameter estimates are obtained by means of an automatic segmentation procedure for careful modeling of relevant pho etic features. The model structure and the training scheme are directed to associate the most acoustically dis criminativesegments of the consonants with a sequence of states in every consonant model. The speech data base consisted of a total of 2, 592 productions of the Spanish Stops /p, t, k, b, d, g/ in intervocalic positions with the vowels /a, i, u/ embedded in VCVCVCV nonsense utterances. The speech data has been produced by two male Argentine Spanish speakers. Phoneme recognition is accomplished finding the state sequence with highest likelihood in an ergodic model formed by the linking of all the context-dependent phoneme models allowing only the phonotactically valid state transi tions.A comparative study of the recognition performance under different degrees of context dependence, and the alternative use of spectral dynamic and energy related pa rametersis presented.

Articulatory and acoustic analysis of consonants differing in intelligibility

Articulatory and acoustic recordings of a single talker producing multiple repetitions of nonsense utterances were made at the University of Wisconsin X‐Ray Microbeam facility. The utterances consisted of the fixed frames “Say a Cawa” and “Say a waCa” with C = {/b/, /d/, /g/}. Each utterance was repeated three times in each of three speaking styles ranging from an exaggerated style intended to be “clear” (i.e., easier for the hard of heating to understand) to a relaxed style intended to be “conversational.” The audio recordings were presented to hearing impaired listeners in a three‐alternative forced‐choice stop consonant identification task to determine the relative intelligibility of the stop consonants. By hypothesis, the clearest consonants would be found in utterances that exhibited the greatest amounts of articulatory and acoustic change. However, this pattern was not found. For utterances with syllable‐initial consonants (e.g., “Say a gawa”), actual intelligibility as estimated by percentage correct stop constant identification was inversely related to both the talkers intent and measures of the extent of articulator motion. Correlational analyses using both articulatory and acoustic measures to predict obtained intelligibility scores will be presented. [Work supported by NIH.]

Hidden Markov models for Spanish stops

In this work, the design and evaluation of the recognition performance of hidden Markov models (HMM) for the intervocalic Spanish stops are described. The proposed HMMs were context dependent in order to account for coarticulatory effects. Continuous probability density functions were used for the output vectors, which included spectral dynamic parameters. Initial model parameter estimates were obtained by means of an ad hoc segmentation procedure for careful modeling of spectral transitions. The speech database consisted of a total of 2592 productions of the Spanish stops/p, t, k, b, d, g/ in intervocalic position with the vowels /a, i, u/ in all combinations. The speech data were produced by two male Argentine speakers. As a step to evaluate the recognition performance with running speech, three different consonants were produced embedded in VCVCVCV nonsense utterances. Results were obtained from variations of the basic models that consisted of a comparative study of the recognition performance under different degrees of context dependence, and the alternative use of spectral transition parameters. In the best condition the average consonant recognition performance, obtained in a speaker‐dependent manner, was 96.61% and 96.53% for each speaker. The errors were concentrated mainly in the voiced stops, with almost perfect recognition obtained for the unvoiced stops.

Durations of Italian Diphthongs and Vowel Clusters

Traditional grammars consider the presence of one segment marked [+high] and the presence of stress (lexical accent) on the contiguous [—high] vowel as the main factors responsible for diphthong realization in Italian; yet, they report that diphthongs also occur in unstressed contiguous vowels. These statements are considered here as the hypotheses to be investigated. Assuming that the gliding process of diphthongization involves shortening, durations of Italian vowels and vowel-like sounds in candidate diphthongs and nondiphthongs (i.e., vowel clusters) occurring in both stressed and unstressed sequences within word boundaries were measured in two different speech materials: (1) nonsense words embedded in sentence frames; (2) meaningful words in sentences. in both contexts, the test words contained pairs of contiguous vowels to be pronounced as a diphthong or as a vowel cluster, according to recognized Italian phonotactic possibilities. Results show the following: (1) in both corpora, segmental duration differences do not effectively characterize the realization of diphthongs and vowel clusters in stressed sequences, while in unstressed sequences there is a strong duration difference between a diphthong and the corresponding vowel cluster, but only in the onglide ([+high/-high]) case. (2) Vowel clusters appear to be articulated differently in meaningful sentences than in nonsense utterances, even when the speaker manages to maintain the same nominal speaking rate. These results will be applied in the text-to-speech synthesis of Italian which is now being developed.

Token‐to‐token variation of tongue‐body vowel targets: The effect of coarticulation

Variation of vowel targets for a midsagittal point on the tongue body of a single speaker of American English was examined. The subject pronounced multiple repetitions of nonsense utterances of the form /bV(C)Vbʌ/, in which the vowels were /i/, /u/, /a/ and the consonant, when present, was /b/ or /h/. Stress was placed on the second syllable. Movements of a single point on the middle of the tongue‐body dorsum were transduced with a new alternating magnetic field movement transducer system. For each vowel, a scatter plot was generated of “target” positions (x,y coordinates) at the moment of minimum tangential velocity. In general, these scatter plots are elongated. When the vowels in the first and second syllables are the same, the scatter in target position is relatively tighter and less elongated than when the vowels are different. When the vowels in the two syllables are different, the long axis passing through the scatter for one vowel is oriented in the direction of the target for the other vowel, providing a “statistically based” demonstration of context dependence of articulatory targets. [Work supported by NIH Grant No. NS04332.]