A cloud-computing platform for developing and evaluating vocal biomarkers based on home audio recordings: Resources for large-scale data processing and analysis

Abstract

This presentation describes the design and construction of a comprehensive cloud-computing platform for implementing and evaluating vocal biomarkers, both for assessing typical development of speech and language and detecting deviations in autism. As part of a prospective longitudinal study of early vocal development (NIH P50 MH100029), whole-day home audio recordings were collected from more than 300 children every month from birth to three years using LENA, providing long-term, large-scale sampling of each child's natural environment. The database contains more than 40 000 h of raw audio data, post-processed acoustic measures, event timing statistics, hand-coded and automated annotations, as well as sociodemographic and clinical metadata, the storage, integration and processing of which has presented significant new challenges. Flexible cloud solutions have been adopted to allow secure cross-platform shared access and referencing of the data. Customized modular “C” code libraries for physical modeling, spectral analysis, temporal event labeling, voice activity detection, speaker diarization and developmental profiling have been developed, based on standard BLAS/LAPACK libraries optimized for device-independent virtual machines, and integrated with an interactive graphical interface for signal analysis and labeling. Resources and specific challenges for large-scale analysis of infant vocalizations will be discussed.