Benchmarking an automated speech recognition pipeline for Spanish-English bilingual preschoolers: Quantifying language input and output in daylong recordings.

Measuring how much of each language bilingual children hear and use in their daily lives requires transcribing long, naturalistic recordings, which is too costly to do by hand at scale. We benchmarked four WhisperX-based automatic speech recognition pipelines, differing in how speech is located and attributed to speakers, against expert transcripts of 168 five-minute clips (14 hours) drawn from daylong home recordings of 14 Spanish--English bilingual preschoolers. Word error rates across pipelines were high (69.5% to 82.0% overall), compared with 35.4% (key child) and 53.9% (adults) between expert coders. Running a child-centered voice type classifier (VTC2) before transcription attributed words to speakers far better than labeling speakers afterwards; we retained this pipeline (WER = 71.4%). Despite word-level errors, it preserved the rank order of clips; automated and manual counts of total and distinct words correlated at r = .83 to .93 for adults and the key child in each language. Agreement rose further at the recording level. Lexical diversity was recovered best: slopes for distinct words (0.79 to 0.88) were closer to one than for total words (0.53 to 0.77). The proportion of Spanish was underestimated (adults: 77% vs. 89%; key child: 70% vs. 90%), largely because WhisperX output short English phrases on brief, unclear audio. Open, inspectable pipelines can yield language-specific measures of bilingual adult input and child output, but precise word counts and language proportions require correction before direct interpretation. These estimates, from one Spanish--English community, will need re-estimating in other populations.