Kelly Davis

Mozilla

Free(ing) Speech: Collection, Validation, and Recognition (Bio)

We’ll discuss the past, present, and future of Mozilla’s Common Voice (Collection & Validation) and Deep Speech (Recognition), two projects that are freeing speech by crowdsourcing collection and validation while opening speech algorithms and models. Particular focus will be placed on communications these projects enable, anything from the transcription of Māori radio shows to bootstrapping Kinyarwanda speech technologies in Kigali.

Satoshi Imaizumi

University of Tokyo Health Sciences

Speech Patterns of Children with Neuro-developmental Disorders (Bio)

Recent neuroscientific studies suggest that individuals with neuro-developmental disorders may have atypical neural connectivity within and between language-related cortical areas, which may result in atypical patterns in their spontaneous speech. We tested this hypothesis by analyzing developmental changes in speech patterns of children and adults who stutter. The effects of phrase syntactic class, length, position in sentence, and manner of articulation on stuttering rate for spontaneous speech recorded from two age groups of Japanese children who stutter. The participants were 14 young children (YC: 2-6 years old) and 14 older children (OC: 7-11 years old). In total, 7406 (YC:3870, OC:3536) spoken phrases were recorded from the participants and classified into fluent versus stuttering phrases. Stuttering rate was calculated as the percentage of stuttering versus total number of phrases. The effects of four independent variables on the stuttering rate were analyzed using ANOVA and a multiple logistic regression model of JMP Pro version 14. The explanatory variables were Group (YC vs. OC), Class (Content vs. Function), Length in Number of morae (Short vs. Long), Phrase Position in sentence (1, 2, 3 or more), and Manner of articulation of phrase-initial consonant (Obstruent vs. Sonorant). Following results were obtained. 1) Phrase length and position significantly affect stuttering rate regardless of age. Longer phrases have higher stuttering rate than shorter ones for both the age groups. The stuttering rate is higher at the head of sentence than at the end of it, but this position effect is smaller for the OC group than the YC group. The effects of Length and Position suggest children who stutter have more difficulty to initiate utterances of longer phrases. 2) The significant effect of Class indicates that content phrases have higher stuttering rate than function phrases, and this effect is large for the YC group and small for the OC group. This may suggest difficulty in construction of long content phrases affects stuttering rate more for younger children. 3) Comparing to the YC group, the OC group uses more long content phrases (Length=5 or more) with less short function phrases (Length=2, 3, 4), and has higher stuttering rate even at end of sentence. This age-dependent variations may be due to the required complexity of language use which increases with age. 4) The YC group shows a significant effect of Manner, that is, their stuttering rate is higher when the initial consonant is stops compared to the others. Manner has no effect for the OC group. This result suggests manner of articulation of sentence-head syllable affects stuttering for the YC group. These results suggest that complexities in linguistic phrase generation and its articulatory difficulties are the main factors to generate atypical speech pattern for children who stutter, possibly reflecting atypical coordination between cortical areas working for message generation and articulatory control. Acknowledgements: The authors appreciate all the participants who made this study possible. This study was approved by Research Ethics Committee of Pref. Univ. Hiroshima, and was supported by JSPS Grants.

Charles Yang

University of Pennsylvania

The Role of the Lexicon in Phonological Acquisition and Change (Bio)

How do children establish abstract phonological representations and processes from the gradient acoustic signal? In this talk, I revisit some classic ideas of how the lexicon provides the impetus for the development of phonology. Specifically, I will present a series of computational models that (a) partition the acoustic space for marking lexical distinctions, (b) establish allophonic alternations to maintain the compactness of the lexicon, and (c) assess the outcome of competing phonemic systems, as in the case of language/dialect contact, as a function of lexical processing cost.