Voice biometrics are deployed in telephone banking, device unlock, and assistant personalisation. All of it rests on one assumption: your voice is stable enough that a machine can keep recognising it. I ran a controlled experiment to find out how much of that assumption survives contact with a person who has a cold, is tired, or is just speaking a little differently than usual.
Four speakers, identical sentences, controlled changes in pitch, phonation and articulation, fixed microphone position and gain. Every utterance projected through three different encoder architectures — ECAPA-TDNN, a ResNet speaker encoder, and WavLM-base-plus-sv.
The result
One speaker sat 0.7 semitones below his normal pitch. That is below the threshold where a listener reliably hears any change at all. His verification score dropped 0.238 across all eight sentences.
Pitch was not what moved him. His harmonics-to-noise ratio fell from 10.63 to 8.39 dB over the same block. The encoder was reacting to phonation — how the vocal folds were vibrating — and the pitch number was just the thing that happened to be easy to measure.
Across the whole run, all 30 speaker × condition × encoder cells were negative, 28 of them unanimous across every utterance.
These systems are not mainly disturbed by shouting or disguise. They are disturbed by someone speaking slightly differently in a way nobody would notice.
Credit where it belongs
The displacement effect itself is not my finding. I had it as my headline until an adversarial prior-art audit — an audit I designed specifically to try to kill my own claim — turned up Hughes et al. (Interspeech 2023), who had established it with six trained phoneticians across seventeen conditions.
I withdrew the claim and reframed the paper as a replication and extension across encoder architectures. The audit is published in full, including the claim it killed.
Two things I could not find in prior work
Jitter beats HNR. Pitch deviation and jitter are independently associated with displacement in all three encoders. Harmonics-to-noise ratio adds nothing once jitter is in the model — it correlates 0.55 with jitter and loses all independent power beside it. HNR is the measure most practitioners reach for, and on its own it will attribute the effect to the wrong thing.
Rough phonation breaks F0 trackers in one direction. 10 of 145 utterances carried impossible pitch values, up to 32.1 semitones — a 6.4× frequency ratio no human produces. All 10 sat in rough phonation, none in modal (Fisher exact p = 2.4e-11). Octave errors push estimates upward and essentially never downward, so this is not noise. It is differential measurement error: a variable that is in the model, mismeasured, in one direction, in one identifiable subset of the data. Those observations carried 2.7× the leverage of the rest and up to 13× the Cook's distance, so the contamination loaded onto one specific regression coefficient.
The fix is cheap: validate your F0 tracker per phonation condition, and publish the validation. A pitch coefficient estimated over pooled phonation is not interpretable without it.
The number I deleted
My pre-registered prediction about pitch direction could be reported at p = 0.013. I am not reporting it. Two defensible ways of correcting the same measurement error give opposite answers:
all data, contaminated up 0.0408 down 0.0416 p = 0.947
drop impossible F0 values only up 0.0846 down 0.0565 p = 0.099
restrict to the clean corpus up 0.0556 down 0.0845 p = 0.013
The third correction also removes an entire corpus, so I cannot separate "cleaned the data" from "changed the sample." I published the instability instead of the significant number.
Everything is open
Manuscript, the pre-registration, the full prior-art audit, per-utterance measurement tables for all three encoders (233 rows each), every analysis script including the influence diagnostics, and the 137 source recordings. There is a provenance manifest mapping all 233 rows back to source recording and time offsets, verified by re-running the segmentation with no encoder loaded.
Every number in the paper reproduces from the CSVs without touching the audio.
Paper and data: 10.5281/zenodo.21921958
Full write-up: How little your voice has to change before a machine stops recognising it
Audio licence: research, benchmarking, evaluation and teaching are permitted; use as machine-learning training data and use for voice cloning are not. The speakers are identifiable adults, three under separate commercial voice contract.
Top comments (0)