Early Diagnosis of Autism through Analysis of Pre-Speech Vocalizations
Keshi Dai · 2007 · SIGACCESS Accessibility and Computing · doi:10.1145/1328567.1328575
Summary
This paper proposes an approach for early diagnosis of autism spectrum disorder (ASD) by computationally analyzing the pre-speech vocalizations of infants aged 6 to 18 months. The author notes that while autism can be reliably diagnosed by age 3, and potentially as early as 12 months, existing screening instruments like CHAT, STAT, and SCQ rely on behavioral observation by professionals and parent recall — subjective methods that primarily detect severe cases and may miss milder presentations. The proposed method leverages two existing technologies: the Early Vocalization Analyzer (EVA), which automatically analyzes infant vocalizations to derive a "vocalization age" and can clinically distinguish at-risk infants from typically developing ones, and the visiBabble system, which processes vocalizations in real time and reinforces syllable-like productions through visual feedback. The technical approach uses acoustic landmark detection based on Liu-Stevens theory, identifying three types of landmarks in infant babbling: glottis markers (vocal fold transitions), sonorant markers (consonantal closures and releases), and burst markers (stop/affricate events). From these landmarks, the system extracts syllable features including syllable rate, syllable number, landmarks per syllable, and syllable duration. The planned study would record younger siblings of children with autism — a population at moderate genetic risk — at Children's Hospital Boston every three months from age one to two years, then retrospectively analyze vocalization features after clinical diagnosis at ages 2.5 to 3 years to identify distinguishing patterns.
Key findings
As a proposal paper, this work does not report experimental results but outlines a technically grounded hypothesis and methodology. The key technical contribution is the identification of specific acoustic features that could serve as early biomarkers for autism: landmark patterns, particular syllable types, anomalous pitch patterns, and vocalization age delay. The system can detect 38 possible syllable patterns from landmark sequences and extract four syllable feature types. The rationale draws on established findings that vocalization age delay correlates with later communication difficulties, and that autism is characterized by impaired social interaction and communication from very early in development. The author references existing evidence that EVA can already distinguish at-risk infants from typically developing peers between 6 and 15 months, suggesting that extending this analysis with autism-specific classification rules is a feasible next step. The planned longitudinal design — recording at-risk siblings prospectively and analyzing retrospectively after diagnosis — represents a methodologically sound approach to identifying pre-diagnostic vocal markers.
Relevance
This paper is significant for accessibility practitioners because it represents an early example of using computational speech analysis as an objective screening tool for developmental disability, moving beyond subjective behavioral observation. If successful, such technology could dramatically lower the age of autism detection, enabling earlier intervention, which research consistently shows leads to better outcomes in communication, motor skills, and cognitive development. The approach is particularly relevant to the broader field of assistive technology as it demonstrates how signal processing and machine learning techniques can be applied to accessibility challenges beyond the traditional domains of screen readers and physical accommodation. The work also raises important considerations around screening ethics, including the balance between early detection benefits and the risks of false positives in vulnerable populations. As a 2007 proposal, the approach anticipated what has since become an active research area in computational behavioral analysis for autism.
Tags: autism spectrum disorder · early diagnosis · speech recognition · pre-speech vocalization · early intervention · assistive technology