On-device human voice removal

A microphone that lives in a household

Audio next to the body is genuinely informative about an animal. Chewing, grooming, movement against fabric and vocalisation are all easier to detect at the neck than across a room. The same proximity means the sensor is also inside every conversation that happens near the animal—kitchens, guests, a television left on, a phone call in the next chair.

The concern this raises is concrete rather than abstract: a household asks whether the cat ate, and in exchange a device retains last night’s discussion. The reason we treat this as a research problem rather than a policy sentence is that the useful signal and the sensitive signal arrive in the same recording, and separating them is a signal-processing question with a storage deadline attached.

Why access control is the wrong boundary

The common design writes the recording as captured and then restricts who may open it. That places the entire protection on later behaviour—permissions, retention jobs, the integrity of every future feature that touches the file. It also means the sensitive material exists, and the question of whether it survives becomes a question about operational discipline.

We prefer a boundary that does not depend on subsequent good conduct. The decision point is moved to the moment before anything durable exists: the incoming payload is completed in memory, speech is removed there, and only the derived recording is written. There is no earlier version on the device to be protected, exported or recovered later, and no temporary artefact standing in for one.

Failing closed without losing the timeline

A privacy step that can fail needs a defined behaviour for failure, and “write the original instead” is not it. When processing cannot complete safely—an unusable sample rate, an internal error, a required component unavailable—the path emits a recording of the same duration carrying no speech, retains the accompanying motion data, and records why.

Preserving duration is what keeps the guarantee compatible with the rest of the system. Inertial data from the collar and video from the room are aligned against these recordings, so a shortened or missing file would silently distort a later reconstruction of the afternoon. To everything downstream, a clip that never contained speech and a clip whose speech was removed are the same kind of object, differing only in the metadata that describes what happened to it.

An experiment at the scale of a lived-in home

A rule of this kind is easy to demonstrate on a handful of clips and hard to trust until it has met a messy corpus. Our development hub had been recording in a real household for months, which gave us an experimental set with the properties we could not synthesise: unequal clip lengths, uneven speech density, conversations at varying distance, and a long tail of recordings captured before the boundary existed at all.

The experiment applied the rule to that whole accumulated set rather than to a curated subset, beginning with a twenty-record trial whose outcomes were inspected individually before the full run was authorised. In the trial, fourteen recordings contained no speech, six had speech removed, and none failed.

Across the full experimental set of 10,674 recordings, 7,841 contained no speech and were left as captured; 2,292 had speech removed; and 541 could not be processed safely and were written as equal-duration recordings without speech, each with its reason attached. A recording captured live during the run had a little over half of its timeline identified as speech and rewritten, with duration and accompanying motion data intact.

The distribution is the useful part of the result. Roughly three quarters of the recordings from an occupied home turned out to contain no human speech at all, which is the case where the method must stay out of the way; the failure count told us that the fail-closed path is not a theoretical branch but one that real conditions reach, and that it therefore has to preserve the timeline rather than drop the clip.

An audit trail that contains no audio

Removing data creates an evidence problem. If nothing records what happened, a privacy guarantee becomes an assertion, and the system cannot distinguish a clip that never contained speech from one whose processing quietly failed.

Each recording therefore carries a compact receipt: what the source measured, what was produced, which outcome applied, and how much of the timeline was affected. The receipt is verifiable against the stored recording and holds nothing from which audio could be reconstructed. Auditing the experimental set afterwards was how we checked the run rather than assumed it—every recording had a matching receipt, every derived file matched its recorded fingerprint, motion data was intact throughout, and no recording that had failed closed was found to contain residual audio.

Scope of the claim

The results above come from an experiment on our own development hardware, and they characterise the method rather than describe a shipped configuration. They are also a statement about what a device persists, not a claim that a home will be free of human sound: other household noise remains, speech detection is imperfect at the margins, and the value of the method depends on removal being aggressive enough to matter while leaving the animal audible.

That tension is the active research question rather than a solved one: a detector tuned to remove more conversation will eventually remove chewing and purring with it, and a recording that no longer contains the animal is useless for the purpose that justified recording at all. The commitment we are willing to publish is narrower and, we think, more useful than a reassurance—the hub does not require a family’s sentences in order to understand an animal, so it should not retain them.