Comment by kzahiri

It isn't bad. On 2,760 Common Voice clips, with age brackets matched on accent, gender and speaker so age is the only thing varying, Whisper large-v3 gets 4.67% WER on speakers in their seventies against 6.53% for twenty-somethings. Substitutions, deletions and insertions all fall with age. Deletions in particular don't rise, which is what you'd see if quiet speech were being dropped.

Since Whisper's decoder is a language model and could plausibly be repairing older speakers' word choices rather than hearing them better, I re-ran the same clips through wav2vec2, which is pure CTC with no decoder and no LM: 10.30% for the sixties against 14.23% for the twenties. Larger effect, same direction. So it's acoustic, not a decoder artifact.

Where age does hurt is turn-taking. Voice agents end your turn after a fixed stretch of silence, typically 500-800ms. At 700ms, 8.0% of twenties utterances contain an internal pause long enough to be misread as end-of-turn, versus 19.7% for the sixties. Older speakers take about twice as many pauses inside a single utterance. WER is structurally blind to this: the words that arrive are transcribed correctly while the speaker gets cut off.

A maintainer corrected me on that and he was right. Mark Backman at Daily/Pipecat pointed out that production stacks don't endpoint on a bare VAD threshold anymore. Pipecat's default is smart-turn, a semantic model that listens to the waveform. So I measured smart-turn v3 on the same clips: the gap halves to +5.9pp and stops excluding zero. The fixed-threshold result describes a real and common configuration, not every configuration, and the README says so above the fold.

Things that mattered methodologically, mostly because they nearly produced wrong answers:

- One Common Voice contributor holds 9,792 clips in a single shard, and about seven people account for half the 60+ audio. Uncapped, a "bracket" is a description of one person. Capped at 25 clips each; all intervals bootstrap over speakers, not clips.

- The corpus's younger contributors skew non-native, so age and accent are entangled. Brackets are matched on the (accent, gender) pair. I also spent a while believing accent explained the whole result, based on a 40-clip pilot; at full sample it moves WER by at most 0.44pp. The wrong claim is corrected in the README rather than deleted.

- My first VAD was manufacturing the headline. A relative-energy threshold called breathy trailing-off speech "silence", and breathiness correlates with age, so the detector's error was correlated with the variable under study. WebRTC VAD disagreed on 36% of the eighties clips. Both are now recorded.

- fp16 on MPS is verified against fp32/CPU rather than assumed, because a quantised Whisper KV cache can take large-v3 from 1.91% WER to 100%.

There's a fourth result in there that's really about a different industry. Several startups now sell daily phone check-ins for older adults claiming to detect cognitive decline from "voice biomarkers". Validating that needs gated clinical corpora, but the within-speaker noise floor doesn't. Pause features vary about 99% within the same speaker, frequently within one sitting. Getting a 10% shift in total pause time above that noise takes roughly 19 days of daily calls per reading, so a "six-week trend" is two or three noisy measurements. Speech rate needs under one call. Pause metrics are the most-cited biomarker and the least usable for per-person drift.

Limits, stated because they're real: Common Voice's older speakers are volunteers who chose to record themselves, so this is healthy aging, not clinical. It's read speech, not conversation. And the disordered-speech follow-up is blocked on corpus licensing — TORGO and UASpeech need signed agreements and their HuggingFace mirrors look like unlicensed redistribution, so I didn't use them. The one openly-licensed option turned out to be fixed 3-second podcast excerpts where 74% start mid-speech, which makes the pause metric meaningless on it. That failure is written up in METHOD.md instead of being quietly dropped.

Everything runs locally, no API key, no spend.