Whisper's WER Drops with Age — Contrary to Assumptions
A new benchmark from Kayvan-Zahiri, published as an HN submission, flips a common assumption: Whisper large-v3 actually transcribes older speakers more accurately. On 2,760 matched Common Voice clips, the word error rate (WER) for speakers in their twenties was 6.53%, but for sixties it dropped to 5.23% and for seventies to 4.67%. The differences are statistically significant (speaker-bootstrapped 95% intervals exclude zero).
The effect isn't an artifact of Whisper's language model. Re-running the same clips through wav2vec2 (pure CTC, no decoder) shows an even larger gap: WER for twenties was 14.23%, dropping to 10.30% for sixties and 10.52% for seventies. This points to acoustics, not a decoder repair.
All error types (substitutions, deletions, insertions) fall with age. Deletions, which would rise if quiet speech were being dropped, stay flat. The authors note the corpus measures healthy aging, not clinical conditions like dysarthria.
The Real Problem: Turn-Taking
While recognition improves, voice agents that endpoint on a fixed silence threshold talk over older speakers. At 700ms, 8.0% of utterances from twenties contain an internal pause long enough to trigger a false cutoff, but 19.7% for sixties and 16.6% for seventies. That's a +11.6pp and +8.5pp difference, both significant.
The mechanism: older speakers take about twice as many internal pauses and spend twice as long in them. The effect plateaus by age 60, and the eighties (analyzed separately due to small sample) show a 5.4x gap at 700ms.
The result survives a re-draw without accent matching: 19.3% and 16.6% for sixties and seventies, nearly identical to the matched numbers.
Semantic Turn Models Close Most of the Gap
Production stacks don't use a fixed threshold. Pipecat's smart-turn v3, a semantic model, reduces the gap: at 700ms, the sixties gap drops from +11.6pp to +5.9pp, and the interval now includes zero. The absolute rate (76-82%) isn't an error rate — many pauses are legitimate clause boundaries — but the between-bracket comparison is interpretable.
The practical takeaway: if you use a fixed VAD threshold, the age gap is real and large. If you use a semantic turn model, most of it goes away. The published smart-turn benchmark doesn't stratify by age, and its training mix leans on synthetic TTS, which pauses differently than an 80-year-old.
Voice Biomarkers: Within-Speaker Noise Is Huge
Products that claim to detect cognitive decline from voice biomarkers need to clear the speaker's own noise. This benchmark measures within-speaker coefficient of variation (CV) across 36 speakers with 40+ clips each:
- Speech rate: ~18% CV
- Utterance duration: ~23% CV
- Number of internal pauses: ~76% CV
- Total pause time: ~96-111% CV
Pause features vary by ~100% within a single speaker. Detecting a 10% change in total pause time at 80% power requires about 758 utterances — roughly 19 daily calls. A "six-week trend" is two or three noisy measurements. Speech rate and duration are more usable.
The Accent Confound: Real but Not Decisive
Common Voice's younger contributors skew non-native, and Whisper is worse on non-native English. The author matched brackets on (accent, gender) to hold confounds constant. The result holds: matched and unmatched samples show the same pattern, with the largest disagreement at 0.44pp. The original pilot suggested otherwise, but that was noise.
Reproduction and Caveats
The benchmark is reproducible with python3 bench/run.py — no API key, no spend. The code is on GitHub.
Caveats: Common Voice's older speakers are tech-comfortable volunteers, not representative of post-discharge patients. Read speech differs from conversational speech, though that cuts against the turn-taking finding being an artifact. The eighties bracket has only 27 speakers, so intervals are wide.
What This Means for Voice Agent Developers
- Don't assume older voices break ASR. Whisper's WER improves with age on this corpus. But don't extrapolate to clinical populations.
- Replace fixed VAD thresholds with semantic turn models. If you're using a fixed 700ms threshold, you're cutting off older callers 2-2.5x more often. Pipecat's smart-turn v3 halves the gap.
- If you build voice biomarkers, account for within-speaker noise. Pause features are nearly 100% CV — a single measurement is meaningless.
Next Steps
Run the benchmark yourself. Then audit your voice agent's endpointing logic. If you use a fixed threshold, switch to a semantic model. If you're building biomarker analysis, collect multiple baseline sessions before making claims.
The author notes the honest next step is a follow-up on disordered-speech corpora. Until then, treat these results as a measure of healthy aging.
