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 real failure that is happening. Measured on 2,760 Common Voice clips, matched on accent, gender, and speaker, the only variable is age, and a second 3,189‑clip draw controls for none of it.

Whisper transcribes older speakers more accurately, not less. The 95% bootstrap interval for WER shows a ‑1.31pp difference for sixties versus twenties, and a ‑1.86pp for seventies, with every error type falling with age. Re‑running the same clips through wav2vec2, a pure acoustic model, reproduces the effect, confirming it is acoustics, not a language‑model repair.

Turn‑taking penalties are large when a voice agent endpoints on a fixed silence threshold. At a 700‑ms cutoff, older speakers are cut off 2‑2.5× more often, with WER gaps of +11.6pp for sixties and +8.5pp MT for seventies. The mechanism is internal pauses; older speakers take twice as many and spend twice as long in them.

A semantic turn model, such as Pipecat’s smart‑turn v3, closes most of the gap. The same 700‑ms cutoff with smart‑turn reduces the WER difference to +5.9pp for sixties and +4.0pp for seventies, and the interval now includes zero, showing that most age‑related penalties vanish when endpointing is based on speech content rather than silence.