Phonemic Context Research Articles

Developing an algorithm to characterize context-based speech patterns as cue production profiles

The surface articulation of an underlying phoneme commonly fluctuates depending on the context. For instance, the associated acoustic cues for a /t/ phoneme produced at the beginning of an utterance may be distinct from the acoustic cues of a /t/ phoneme that is preceded and followed by a vowel. Thus, for an individual speaker or speaker group, we can investigate what specific phonemic contexts result in the associated acoustic cues to be produced. An algorithm has been developed which matches a target phoneme to its corresponding acoustic cues. For a given database, we can recover all of the contexts in which a phoneme was produced and tabulate the various acoustic cue production patterns that arose. Finally, we label these patterns by their speech production type (e.g., standard production, flapping, etc.) and produce an output in the form of a cue production profile. As a result, this profile would allow for the association of a set of acoustic cues in specific phonemic contexts for an individual speaker or speaker group, such as speakers from a specific dialect region. The algorithm can also be expanded to include prosodic cues to account for how stress and intonation can affect phoneme production.

Feasibility of using ultrasound visual biofeedback to treat persistent speech sound disorders in children with cleft palate- a case series

ABSTRACT The current study aimed to assess the effectiveness of incorporating ultrasound visual biofeedback (UVB) into a treatment programme addressing persistent speech sound disorders linked to cleft palate in children who have been unresponsive to traditional therapy approaches. Materials and Methods. A single-subject multiple baseline experiment was conducted with five children aged 6:5–13:5 over a period of 16 therapy sessions. Treatment focused on providing cues from real-time ultrasound images to assist children in modifying their tongue movements. Probe data were collected before, mid, and post-treatment to assess target consonant accuracy for 50 untreated words. The results of the statistical analysis suggested participants showed a significant increase in percent target consonant accuracy as a result of intervention using UVB. Although most of the participants exhibited progress in generalising learned phonemes to untreated words, some did not show improvement in gaining generalisation from treated phonemic contexts to those untreated ones. When traditional methods fail to yield significant progress, incorporating ultrasound biofeedback into the treatment programme emerges as a viable option to enhance sound accuracy in children with persistent speech sound disorders resulting from cleft palate.

Covariants of Gemination in Eastern Andalusian Spanish: /t/ following Underlying /s/, /k/, /p/ and /ks/

In Eastern Andalusian Spanish, consonants are deleted in syllable-final position, triggering regular gemination of a following consonant, even across word boundaries. This paper investigates five underlying phonemic contexts involving /t/, including singleton /t/ and four different underlying /C+t/ sequences that typically surface as [t:], by analyzing how durational and formant differences vary depending on the presence and identity of the preceding underlying consonant. Following the acoustic and statistical analyses of 444 instances of /ˈeta/, /ˈesta/, /ˈekta/, /ˈepta/ and /ˈeksta/, a Discriminant Function Analysis shows that differences in the total duration of /t/ and in the duration of the closure of /t/ are the strongest cues to distinguishing singletons from geminated consonants, with 91.9% and 90.6% accurate classifications, respectively. Cues indicating which specific consonants have been deleted before /t/ are much less robust and more varied in nature. It is unclear, however, whether this outcome is due to different compensation strategies in each case or whether they are affected by some kind of underlying coarticulatory effect. Given that gemination in this language variety is the result of regular /C1C2/ to [Cː] assimilation, and that its underlying phonemic status has not been demonstrated, Eastern Andalusian Spanish is unusual amongst languages studied with respect to gemination, making this study typologically interesting.

The influence of phoneme contexts on adaptation in vowel‐evoked envelope following responses

Repeated stimulus presentation leads to neural adaptation and consequent amplitude reduction in vowel‐evoked envelope following responses (EFRs)—a response that reflects neural activity phase‐locked to envelope periodicity. EFRs are elicited by vowels presented in isolation or in the context of other phonemes such as consonants in syllables. While context phonemes could exert some forward influence on vowel‐evoked EFRs, they may reduce the degree of adaptation. Here, we evaluated whether the properties of context phonemes between consecutive vowel stimuli influence adaptation. EFRs were elicited by the low‐frequency first formant (resolved harmonics) and middle‐to‐high‐frequency second and higher formants (unresolved harmonics) of a male‐spoken /i/ when the presence, number and predictability of context phonemes (/s/, /a/, /∫/ and /u/) between vowel repetitions varied. Monitored over four iterations of /i/, adaptation was evident only for EFRs elicited by the unresolved harmonics. EFRs elicited by the unresolved harmonics decreased in amplitude by ~16–20 nV (10%–17%) after the first presentation of /i/ and remained stable thereafter. EFR adaptation was reduced by the presence of a context phoneme, but the reduction did not change with their number or predictability. The presence of a context phoneme, however, attenuated EFRs by a degree similar to that caused by adaptation (~21–23 nV). Such a trade‐off in the short‐ and long‐term influence of context phonemes suggests that the benefit of interleaving EFR‐eliciting vowels with other context phonemes depends on whether the use of consonant‐vowel syllables is critical to improve the validity of EFR applications.

Sensitivity of Vowel-Evoked Envelope Following Responses to Spectra and Level of Preceding Phoneme Context.

Vowel-evoked envelope following responses (EFRs) could be a useful noninvasive tool for evaluating neural activity phase-locked to the fundamental frequency of voice (f0). Vowel-evoked EFRs are often elicited by vowels in consonant-vowel syllables or words. Considering neural activity is susceptible to temporal masking, EFR characteristics elicited by the same vowel may vary with the features of the preceding phoneme. To this end, the objective of the present study was to evaluate the influence of the spectral and level characteristics of the preceding phoneme context on vowel-evoked EFRs. EFRs were elicited by a male-spoken /i/ (stimulus; duration = 350 msec), modified to elicit two EFRs, one from the region of the first formant (F1) and one from the second and higher formants (F2+). The stimulus, presented at 65 dB SPL, was preceded by one of the four contexts: /∫/, /m/, /i/ or a silent gap of duration equal to that of the stimulus. The level of the context phonemes was either 50 or 80 dB SPL, 15 dB lower and higher than the level of the stimulus /i/. In a control condition, EFRs to the stimulus /i/ were elicited in isolation without any preceding phoneme contexts. The stimulus and the contexts were presented monaurally to a randomly chosen test ear in 21 young adults with normal hearing. EFRs were recorded using single-channel electroencephalogram between the vertex and the nape. A repeated measures analysis of variance indicated a significant three-way interaction between context type (/∫/, /i/, /m/, silent gap), level (50, 80 dB SPL), and EFR-eliciting formant (F1, F2+). Post hoc analyses indicated no influence of the preceding phoneme context on F1-elicited EFRs. Relative to a silent gap as the preceding context, F2+-elicited EFRs were attenuated by /∫/ and /m/ presented at 50 and 80 dB SPL, as well as by /i/ presented at 80 dB SPL. The average attenuation ranged from 14.9 to 27.9 nV. When the context phonemes were presented at matched levels of 50 or 80 dB SPL, F2+-elicited EFRs were most often attenuated when preceded by /∫/. At 80 dB SPL, relative to the silent preceding gap, the average attenuation was 15.7 nV, and at 50 dB SPL, relative to the preceding context phoneme /i/, the average attenuation was 17.2 nV. EFRs elicited by the second and higher formants of /i/ are sensitive to the spectral and level characteristics of the preceding phoneme context. Such sensitivity, measured as an attenuation in the present study, may influence the comparison of EFRs elicited by the same vowel in different consonant-vowel syllables or words. However, the degree of attenuation with realistic context levels exceeded the minimum measurable change only 12% of the time. Although the impact of the preceding context is statistically significant, it is likely to be clinically insignificant a majority of the time.

Do adults produce phonetic variants of /t/ less often in speech to children?

The surface phonetic details of an utterance affect how ‘native’a speaker sounds. However, studies have shown that children’s acquisition of context-appropriate variation (sometimes called allophones) is late. This study’s goal was to understand how caregivers use phonetic variation in the production of American English /t/ in child-directed speech (CDS), compared to in adult-directed speech (ADS). We hypothesized that mothers modify their input to children in order to produce more limited variation in CDS than in ADS, to potentially assist children in the development of contrastive phonemic categories. To this end, we recorded eight mothers of children under the age of 2 years in both ADS and CDS conditions. Results reveal that CDS contains significantly more canonical cues to /t/ than ADS does, and fewer non-canonical cue patterns, including fewer unreleased tokens and fewer glottalized tokens in utterance-medial position. Also, we found larger aspiration duration differences in CDS between aspirated singleton [th] vs. unaspirated [t] in /st/ contexts, suggesting that mothers exaggerate this cue to the phonemic context in which the /t/ occurs. Overall, the findings suggest that CDS more clearly signals the phonemic category, which could in turn assist children learn the relationship between the underlying and surface forms.

Effect of Clear Speech on the Duration of Silent Intervals at Syntactic and Phonemic Boundaries in the Speech of Individuals With Parkinson Disease.

Purpose The purpose of the current investigation was to determine the extent to which individuals with and without Parkinson disease (PD) modified silent interval durations when using a clear speaking style. Method Ten individuals with idiopathic PD and 10 older adult control speakers produced a reading passage using both habitual and clear speaking styles. Silent intervals lasting 15 ms and longer were identified and extracted. Each silent interval was categorized according to the surrounding syntactic context of the reading passage. In addition, voiceless stop gaps that occurred within words, phrases, or clauses were categorized by the preceding phonemic context. Results Statistical analyses revealed that all participants increased silent interval duration with a clear speaking style at inter-sentence and intra-sentence syntactic boundaries. Compared to controls, individuals with PD exhibited significantly less increase in silent interval durations at these syntactic boundaries. Control speakers also increased silent stop gap durations in the clear speaking style regardless of preceding phonemic context. Individuals with PD, however, only increased stop gap duration when the silent interval was preceded by a sonorant. Conclusion These findings suggest that speakers with PD exhibit less clarity-related increase in silent interval duration than control speakers. In addition, speakers with PD exhibited significant increases in silent interval duration that coincided with syntactic boundaries of the reading passage but little to no clarity-related modulation of stop gap intervals. Therefore, these data suggest that speakers with PD exhibited changes in silent interval durations that were more so associated with modulation of speech prosody than articulation when using a clearer speaking style.

Effects of Biofeedback on Control and Generalization of Nasalization in Typical Speakers.

The purpose of this study was to determine the effects of biofeedback on control of nasalization in individuals with typical speech. Forty-eight individuals with typical speech attempted to increase and decrease vowel nasalization. During training, stimuli consisted of consonant-vowel-consonant (CVC) tokens with the center vowels /a/ or /i/ in either a nasal or nonnasal phonemic context (e.g., /mim/ vs. /bib/), depending on the participant's training group. Half of the participants had access to augmentative visual feedback during training, which was based on a less-invasive acoustic, accelerometric measure of vowel nasalization-the Horii oral-nasal coupling (HONC) score. During pre- and posttraining assessments, acoustically based nasalance was also measured from the center vowels /a/, /i/, /æ/, and /u/ of CVCs in both nasal and nonnasal contexts. Linear regressions indicated that both phonemic contexts (nasal or nonnasal) and the presence of augmentative visual feedback during training were significant predictors for changes in nasalance scores from pre- to posttraining. Participants were able to change the nasalization of their speech following a training period with HONC biofeedback. Future work is necessary to examine the effect of such training in individuals with velopharyngeal dysfunction.

The Effect of Age and Type of Noise on Speech Perception under Conditions of Changing Context and Noise Levels

Objective: Everyday life includes fluctuating noise levels, resulting in continuously changing speech intelligibility. The study aims were: (1) to quantify the amount of decrease in age-related speech perception, as a result of increasing noise level, and (2) to test the effect of age on context usage at the word level (smaller amount of contextual cues). Patients and Methods: A total of 24 young adults (age 20-30 years) and 20 older adults (age 60-75 years) were tested. Meaningful and nonsense one-syllable consonant-vowel-consonant words were presented with the background noise types of speech noise (SpN), babble noise (BN), and white noise (WN), with a signal-to-noise ratio (SNR) of 0 and -5 dB. Results: Older adults had lower accuracy in SNR = 0, with WN being the most difficult condition for all participants. Measuring the change in speech perception when SNR decreased showed a reduction of 18.6-61.5% in intelligibility, with age effect only for BN. Both young and older adults used less phonemic context with WN, as compared to other conditions. Conclusion: Older adults are more affected by an increasing noise level of fluctuating informational noise as compared to steady-state noise. They also use less contextual cues when perceiving monosyllabic words. Further studies should take into consideration that when presenting the stimulus differently (change in noise level, less contextual cues), other perceptual and cognitive processes are involved.

Cortical encoding of phonemic context during word production.

Brain-computer interfaces that directly decode speech could restore communication to locked-in individuals. However, decoding speech from brain signals still faces many challenges. We investigated decoding of phonemes - the smallest separable parts of speech - from ECoG signals during word production. We expanded on previous efforts to identify specific phoneme by identifying phonemes by where in the word they were formed. We evaluated how the context of phonemes in words affects classification results using linear discriminant analysis. The decoding accuracy of our linear classifier indicated the degree to which the context of a phoneme can be determined from the cortical signal significantly greater than chance. Further, we identified the spectrotemporal features that contributed most to successful decoding of phonemic classes. Finally, we discuss how this can augment speech decoding for neural interfaces.

Processing Nasals with and without Consecutive Context Phonemes: Evidence from Explicit Categorization and the N100

With neurophysiological (N100) and explicit behavioral measures (two-alternative forced-choice categorization), we investigated how the processing of nasal segments of German is affected by following context phonemes and their place of articulation. We investigated pre-lexical processing, with speech stimuli excised from naturally spoken utterances. Participants heard nasals (/n/, /m/, and place-assimilated /n′/), both with and without a subsequent context phoneme. Context phonemes were voiced or voiceless, and either shared or did not share their place of articulation with the nasals. The explicit forced-choice categorization of the isolated nasals showed /n′/ to be in-between the clear categorizations for /n/ and /m/. In early, implicit processing, /m/ had a significantly higher N100 amplitude than both /n/ and /n′/, with, most importantly, no difference between the latter two. When presented in context (e.g., /nb/, /mt/), explicit categorizations were affected by both the nasal and the context phoneme: a consecutive labial led to more M-categorizations, a following alveolar to more N-categorizations. The early processing of the nasal/+context stimuli in the N100 showed strong effects of context, modulated by the type of preceding nasal. Crucially, the context effects on assimilated nasals /n′/ were clearly different to effects on /m/, and indistinguishable from effects on /n/. The grouping of the isolated nasals in the N100 replicates previous findings, using magnetoencephalography and a different set of stimuli. Importantly, the same grouping was observed in the nasal/+context stimuli. Most models that deal with assimilation are either challenged by the mere existence of phonemic context effects, and/or use mechanisms that rely on lexical information. Our results support the existence, and early activation, of pre-lexical categories for phonemic segments. We suggest that due to experience with assimilation, specific speech-sound categories are flexible enough to accept (or even ignore) inappropriate place cues, in particular when the appropriate place information is still present in the signal.

Combining Long Short-Term Memory and Dynamic Bayesian Networks for Incremental Emotion-Sensitive Artificial Listening

The automatic estimation of human affect from the speech signal is an important step towards making virtual agents more natural and human-like. In this paper, we present a novel technique for incremental recognition of the user's emotional state as it is applied in a sensitive artificial listener (SAL) system designed for socially competent human-machine communication. Our method is capable of using acoustic, linguistic, as well as long-range contextual information in order to continuously predict the current quadrant in a two-dimensional emotional space spanned by the dimensions valence and activation. The main system components are a hierarchical dynamic Bayesian network (DBN) for detecting linguistic keyword features and long short-term memory (LSTM) recurrent neural networks which model phoneme context and emotional history to predict the affective state of the user. Experimental evaluations on the SAL corpus of non-prototypical real-life emotional speech data consider a number of variants of our recognition framework: continuous emotion estimation from low-level feature frames is evaluated as a new alternative to the common approach of computing statistical functionals of given speech turns. Further performance gains are achieved by discriminatively training LSTM networks and by using bidirectional context information, leading to a quadrant prediction F1-measure of up to 51.3 %, which is only 7.6 % below the average inter-labeler consistency.

A Frame-Based Context-Dependent Acoustic Modeling for Speech Recognition

We propose a novel acoustic model for speech recognition, named FCD (Frame-based Context Dependent) model. It can obtain a probability distribution by using a top-down clustering technique to simultaneously consider the local frame position in phoneme, phoneme duration, and phoneme context. The model topology is derived from connecting left-to-right HMM models without self-loop transition for each phoneme duration. Because the FCD model can change the probability distribution into a sequence corresponding with one phoneme duration, it can has the ability to generate a smooth trajectory of speech feature vector. We also performed an experiment to evaluate the performance of speech recognition for the model. In the experiment, 132 questions for frame position, 66 questions for phoneme duration and 134 questions for phoneme context were used to train the sub-phoneme FCD model. In order to compare the performance, left-to-right HMM and two types of HSMM models with almost same number of states were also trained. As a result, 18% of relative improvement of tri-phone accuracy was achieved by the FCD model.

Is phonological context always used to recognize variant forms in spoken word recognition? The role of variant frequency and context distribution.

Several mechanisms have been proposed to account for how listeners accommodate regular phonological variation in connected speech. Using a corpus analysis and 5 cross-modal priming experiments, the authors investigate phonological variant recognition for the American English word-final flap. The corpus analysis showed that the flap variant occurs relatively frequently compared with the citation form [t] variant and is only probabilistically constrained by prosodic and phonemic context. The experienced distribution of the flap production is reflected in lexical processing: 4 cross-modal priming experiments demonstrated that lexical activation is not influenced by contextual constraints (inappropriate phrase boundary or phonemic contexts). A 2nd finding was a smaller priming effect for the less frequent flap as compared with the more frequent [t] variant. The contrasts between these findings for the flap and other context conditioned variants are discussed in terms of their implications for models of phonological variation recognition and in terms of the role of language experience.

The phonotactic influence on the perception of a consonant cluster /pt/ by native‐English and native‐Polish listeners: A behavioral and ERP study.

The perception of the cluster /pt/ in word onset in native‐Polish and native‐English adult listeners was compared using behavioral and event‐related potential (ERP) measures to explore the language‐specific phonotactic influence on speech perception. Acoustic‐phonetic features of phonemes vary with context and phonotactic rules specify all allowable phoneme contexts. Both native‐Polish and native‐English groups are exposed to the cluster /pt/ in their language but only one group, native‐Polish, experience the cluster in the context examined, word onset. Word onset consonant cluster /st/ is used as a control. Participants were asked to determine whether they heard the second word of a pair as consisting of two syllables or three (e.g., pteva versus peteva). Behavioral results revealed that Polish listeners perceive the /pt/ cluster with greater accuracy than English listeners. ERPs revealed late components that differ for the two language groups reflecting linguistically‐relevant aspects of the /pt/ contrast and early components that are essentially the same in the two language groups reflecting physical‐acoustic differences. ERP results suggest the /st/ contrast to be more salient than the /pt/ contrast consistent with the higher behavioral accuracy for the /st/ cluster. These findings support the suggestion that psychoacoustic salience of acoustic‐phonetic features influences speech perception.

Effects of Place of Articulation Changes on Auditory Neural Activity: A Magnetoencephalography Study

In casual speech, phonemic segments often assimilate such that they adopt features from adjacent segments, a typical feature being their place of articulation within the vocal tract (e.g., labial, coronal, velar). Place assimilation (e.g., from coronal /n/ to labial /m/: rainbow→*raimbow) alters the surface form of words. Listeners' ability to perceptually compensate for such changes seems to depend on the phonemic context, on whether the adjacent segment (e.g., the /b/ in “rainbow”) invites the particular change. Also, some assimilations occur frequently (e.g., /n/→/m/), others are rare (e.g., /m/→/n/). We investigated the effects of place assimilation, its contextual dependency, and its frequency on the strength of auditory evoked mismatch negativity (MMN) responses, using pseudowords. Results from magnetoencephalography (MEG) revealed that the MMN was modulated both by the frequency and contextual appropriateness of assimilations.

Nasal coarticulation in clear speech.

This study tests whether speakers, when trying to speak clearly, employ variable enhancement strategies as a function of phonetic environment. Using aerodynamic and acoustical methods, this study examines the effects of phonemic context and speaking modality and on the spatial and temporal extent of anticipatory nasal coarticulation in English. Target words are English (C)VNCvoiced (e.g., bend) and (C)VNCvoiceless (e.g., bent) words spoken in either clear or citation speech modes. In order to enhance the percept of /n/ in clear speech, speakers increase the duration of the nasal consonant in CVNCvoiced words but marginally increase, or even decrease, /n/ duration in CVNCvoiceless words. While highly variable, airflow results suggest little difference on anticipatory nasalization as a function of speech mode. These results argue against models predicting a global reduction in coarticulation in clear speech.

Phonetic discrimination is affected by native language phonology.

English and Japanese speakers were tested on discrimination of phonetic contrasts that vary in their phonemic status according to phonological context. The presence versus absence of /u/ in English stimulus pair ebzo‐ebuzo is nonphonemic in Japanese due to phonological syllable structure constraints, but in word‐initial position (stimulus pair zobe‐uzobe the same contrast is phonemic. Similarly, the contrast between Japanese /d/ versus /r/ is nonphonemic in the foot‐medial position in American English due to an allophonic flapping process, but in word‐initial position the two sounds are perceived as the English phonemic contrast /d/ versus /l/ (Japanese stimulus pairs gudo‐guro and dogu‐rogu. Participants were tested in a four‐token oddball task on all four stimulus sets in a blocked design. Each participant group (Japanese and English) showed significantly poorer discrimination of the non‐native contrast when it appeared in the phonological context in which it is nonphonemic in their native language (p’s < 0.001). This result was equally robust even for stimuli in which the contrastive portion of the stimulus (zo‐uzo, do‐ro) was spliced from the nonphonemic context and pasted into the phonemic context. These data indicate that phonological context plays an integral role in the perception of phonetic signals.

The impact of strong assimilation on the perception of connected speech.

Models of compensation for phonological variation in spoken word recognition differ in their ability to accommodate complete assimilatory alternations (such as run assimilating fully to rum in the context of a quick run picks you up). Two experiments addressed whether such complete changes can be observed in casual speech, and if so, whether they trigger perceptual compensation. Experiment 1 used recordings of naive speakers and found that the presence of following context supporting place assimilation led to an increase in miscommunication rate when listeners were asked to identify the potentially assimilated words. This result was also obtained when trained phoneticians gave their considered judgments of a subset of the stimuli. Experiment 2 examined the extent to which words articulated correctly by naive speakers (e.g., rum) would be perceived as assimilated and found that compensation for assimilation in these stimuli depended on the type of following phonemic context and the semantic fit with the preceding sentence. These results suggest that place assimilation does involve complete alternations and that the perceptual system can compensate for them in certain circumstances.

Morpheme-Based Modeling of Pronunciation Variation for Large Vocabulary Continuous Speech Recognition in Korean

This paper describes a morpheme-based pronunciation model that is especially useful to develop the pronunciation lexicon for Large Vocabulary Continuous Speech Recognition (LVCSR) in Korean. To address pronunciation variation in Korean, we analyze phonological rules based on phonemic contexts together with morphological category and morpheme boundary information. Since the same phoneme sequences can be pronounced in different ways at across morpheme boundary, incorporating morphological environment is required to manipulate pronunciation variation modeling. We implement a rule-based pronunciation variants generator to produce a pronunciation lexicon with context-dependent multiple variants. At the lexical level, we apply an explicit modeling of pronunciation variation to add pronunciation variants at across morphemes as well as within morpheme into the pronunciation lexicon. At the acoustic level, we train the phone models with re-labeled transcriptions through forced alignment using context-dependent pronunciation lexicon. The proposed pronunciation lexicon offers the potential benefit for both training and decoding of a LVCSR system. Subsequently, we perform the speech recognition experiment on read speech task with 34K-morpheme vocabulary. Experiment confirms that improved performance is achieved by pronunciation variation modeling based on morpho-phonological analysis.