Diarization in Maximally Ecological Recordings: Data from Tsimane Children

Abstract

Daylong recordings may be the most naturalistic and least invasive way to collect speech data, sampling all potential language use contexts, with a device that is unobtrusive enough to have little effect on people’s behaviors. As a result, this technology is relevant for studying diverse languages, including understudied languages in remote settings – provided we can apply effective unsupervised analyses procedures. In this paper, we analyze in detail results from applying an open source package (DiViMe) and a proprietary alternative (LENA ), onto clips periodically sampled from daylong recorders worn by Tsimane children of the Bolivian Amazon (age range: 6-68 months; recording time/child range: 4-22h). Detailed analyses showed the open source package fared no worse than the proprietary alternative. However, performance was overall rather dismal. We suggest promising directions for improvements based on analyses of variation in performance within our corpus.