Audio watermarking

An audio watermark is a signal embedded in synthesized speech that listeners cannot hear but software can detect — a statement that the audio is synthetic, carried inside the audio itself.

100%

of samples marked at generation

Why the waveform, not the file

A cloning system good enough to be useful is good enough to be misused, and metadata is the wrong defense — file headers are gone after the first transcode. A watermark travels in the waveform itself, so the claim “this is synthetic” rides wherever the audio goes: through the phone network, through a re-encode, into a case file.

What a watermark is not

Not audible — it does not change what a listener hears.
Not speaker identification: it marks audio as synthetic, it does not say whose voice was cloned.
Not optional here — it is applied at generation on every serving tier, part of synthesis rather than a post-process a configuration can skip.
Not the whole provenance story: consent, training-set policy, and per-key accounting carry the rest — collected on the provenance sheet.

Related sheets.

capability

never

Provenance and privacy, applied at generation

The policy surface of the engine: an inaudible watermark on every sample, reference audio that never trains, and per-key audit trails at GET /v1/usage.

glossary

10 s

Instant voice cloning

Instant cloning builds a voice from a ten-second reference at request time — no training job. What zero-shot means, and how identity survives translation.

glossary

~1.5 MB

Voice cloning reference audio

The reference is the clip a zero-shot clone is built from: about ten seconds of one clean, consenting speaker, a ~1.5 MB cap, fingerprinted and cached.

Every term on this page is measurable on a live call — watch the readout while your own script synthesizes.

See the measured figures live