Motoric Mechanisms for the Emergence of Non-local Phonological Patterns

Deep learning assessment of syllable affiliation of intervocalic consonants

In English, a sentence like "He made out our intentions." could be misperceived as "He may doubt our intentions." because the coda /d/ sounds like it has become the onset of the next syllable. The nature and occurrence condition of this resyllabification phenomenon are unclear, however. Previous empirical studies mainly relied on listener judgment, limited acoustic evidence, such as voice onset time, or average formant values to determine the occurrence of resyllabification. This study tested the hypothesis that resyllabification is a coarticulatory reorganisation that realigns the coda consonant with the vowel of the next syllable. Deep learning in conjunction with dynamic time warping (DTW) was used to assess syllable affiliation of intervocalic consonants. The results suggest that convolutional neural network- and recurrent neural network-based models can detect cases of resyllabification using Mel-frequency spectrograms. DTW analysis shows that neural network inferred resyllabified sequences are acoustically more similar to their onset counterparts than their canonical productions. A binary classifier further suggests that, similar to the genuine onsets, the inferred resyllabified coda consonants are coarticulated with the following vowel. These results are interpreted with an account of resyllabification as a speech-rate-dependent coarticulatory reorganisation mechanism in speech.

An informal logic of feedback-based temporal control

A conceptual framework and mathematical model of the control of articulatory timing are presented, in which feedback systems play a fundamental role. The model applies both to relatively small timescales, such as within syllables, and to relatively large timescales, such as multi-phrase utterances. A crucial distinction is drawn between internal/predictive feedback and external/sensory feedback. It is argued that speakers modulate attention to feedback to speed up and slow down speech. A number of theoretical implications of the framework are discussed, including consequences for the understanding of syllable structure and prosodic phrase organization.

Localizing category-related information in speech with multi-scale analyses

Measurements of the physical outputs of speech-vocal tract geometry and acoustic energy-are high-dimensional, but linguistic theories posit a low-dimensional set of categories such as phonemes and phrase types. How can it be determined when and where in high-dimensional articulatory and acoustic signals there is information related to theoretical categories? For a variety of reasons, it is problematic to directly quantify mutual information between hypothesized categories and signals. To address this issue, a multi-scale analysis method is proposed for localizing category-related information in an ensemble of speech signals using machine learning algorithms. By analyzing how classification accuracy on unseen data varies as the temporal extent of training input is systematically restricted, inferences can be drawn regarding the temporal distribution of category-related information. The method can also be used to investigate redundancy between subsets of signal dimensions. Two types of theoretical categories are examined in this paper: phonemic/gestural categories and syntactic relative clause categories. Moreover, two different machine learning algorithms were examined: linear discriminant analysis and neural networks with long short-term memory units. Both algorithms detected category-related information earlier and later in signals than would be expected given standard theoretical assumptions about when linguistic categories should influence speech. The neural network algorithm was able to identify category-related information to a greater extent than the discriminant analyses.

Predicting One's Turn With Both Body and Mind: Anticipatory Speech Postures During Dyadic Conversation

In natural conversation, turns are handed off quickly, with the mean downtime commonly ranging from 7 to 423 ms. To achieve this, speakers plan their upcoming speech as their partner’s turn unfolds, holding the audible utterance in abeyance until socially appropriate. The role played by prediction is debated, with some researchers claiming that speakers predict upcoming speech opportunities, and others claiming that speakers wait for detection of turn-final cues. The dynamics of articulatory triggering may speak to this debate. It is often assumed that the prepared utterance is held in a response buffer and then initiated all at once. This assumption is consistent with standard phonetic models in which articulatory actions must follow tightly prescribed patterns of coordination. This assumption has recently been challenged by single-word production experiments in which participants partly positioned their articulators to anticipate upcoming utterances, long before starting the acoustic response. The present study considered whether similar anticipatory postures arise when speakers in conversation await their next opportunity to speak. We analyzed a pre-existing audiovisual database of dyads engaging in unstructured conversation. Video motion tracking was used to determine speakers’ lip areas over time. When utterance-initial syllables began with labial consonants or included rounded vowels, speakers produced distinctly smaller lip areas (compared to other utterances), prior to audible speech. This effect was moderated by the number of words in the upcoming utterance; postures arose up to 3,000 ms before acoustic onset for short utterances of 1–3 words. We discuss the implications for models of conversation and phonetic control.

Cascading activation in phonological planning and articulation: Evidence from spontaneous speech errors

Speaking involves both retrieving the sounds of a word (phonological planning) and realizing these selected sounds in fluid speech (articulation). Recent phonetic research on speech errors has argued that multiple candidate sounds in phonological planning can influence articulation because the pronunciation of mis-selected error sounds is slightly skewed towards unselected target sounds. Yet research to date has only examined these phonetic distortions in experimentally-elicited errors, leaving doubt as to whether they reflect tendencies in spontaneous speech. Here, we analyzed the pronunciation of speech errors of English-speaking adults in natural conversations relative to matched correct words by the same speakers, and found the conjectured phonetic distortions. Comparison of these data with a larger set of experimentally-elicited errors failed to reveal significant differences between the two types of errors. These findings provide ecologically-valid data supporting models that allow for information about multiple planning representations to simultaneously influence speech articulation.

The Morphogenesis of Speech Gestures: From Local Computations to Global Patterns

A subtle property of speech gestures is the fact that they are spatially and temporally extended, meaning that phonological contrasts are expressed using spatially extended constrictions, and have a finite duration. This paper shows how this spatiotemporal particulation of the vocal tract, for the purpose of linguistic signaling, comes about. It is argued that local uniform computations among topographically organized microscopic units that either constrict or relax individual points of the vocal tract yield the global spatiotemporal macroscopic structures we call constrictions, the locus of phonological contrast. The dynamical process is a morphogenetic one, based on the Turing and Hopf patterns of mathematical physics and biology. It is shown that reaction-diffusion equations, which are introduced in a tutorial mathematical style, with simultaneous Turing and Hopf patterns predict the spatiotemporal particulation, as well as concrete properties of speech gestures, namely the pivoting of constrictions, as well as the intermediate value of proportional time to peak velocity, which is well-studied and observed. The goal of the paper is to contribute to Bernstein's program of understanding motor processes as the emergence of low degree of freedom descriptions from high degree of freedom systems by actually pointing to specific, predictive, dynamics that yield speech gestures from a reaction-diffusion morphogenetic process.