ASR Age Gap - Benchmarking speech recognition and turn-taking across speaker age
Show HN: Whisper transcribes 70-year-olds more accurately than 20-year-olds
ASR Age Gap is an open-source benchmark that challenges the assumption that speech recognition degrades with speaker age. Using matched Common Voice clips, it shows that Whisper and wav2vec2 transcribe older speakers more accurately than younger ones. However, it reveals a critical issue: fixed-threshold turn-taking systems cut off older speakers 2-5x more often due to their longer internal pauses. A semantic turn model can close most of this gap. The project provides reproducible benchmarks, controls for accent and recording quality, and offers insights for building age-inclusive voice agents.
Voice agents are being pointed at elderly callers, and the assumed risk is that speech recognition will not hear them. That assumption is wrong, and it is hiding the failure that is actually happening.