Audio — Vigil
Overview
Vigil is our zero-bit audio watermarking model. It answers one question: does this audio carry our watermark, yes or no?
Use Vigil when you need provenance, not a payload. Typical use cases:
- AI-generation attribution ("was this produced by our TTS?")
- "Is this ours?" content verification
- Any workflow where a robust yes/no decision is enough
Why Vigil is different:
- Extremely robust. Because there's no payload to recover, Vigil can tolerate attacks that defeat most multibit schemes — including over-the-air recapture in non-lab environments.
- Tiny and fast. The detector is a small model designed for low-latency inference, cheap to run at scale.
Imperceptibility
The watermark sits below the perceptual threshold. In a blind A/B test, listeners cannot reliably distinguish the watermarked signal from the original.
Listen for yourself
Select an example to compare the original audio with its Vigil-watermarked version:
Spectrograms
Side-by-side spectrograms of the same clip. The watermarked plot is visually near-identical to the original, demonstrating how little energy is added.
Format & Configuration
| Property | Supported |
|---|---|
| Sample rates | 16, 22.05, 24, 32, 44.1, 48, 96 kHz |
| Channels | Mono, stereo |
| Typical bit depths | 16-bit PCM, 24-bit PCM, 32-bit float |
Robustness
Robustness is Vigil's headline feature. The detector is designed to hold up through real-world transformations that typically break watermarking schemes.
Supported attack classes
| Attack class | Example transformations |
|---|---|
| Lossy compression | MP3 (down to 64 kbps), AAC, Opus |
| Resampling | Arbitrary round-trips (e.g. 48 → 16 → 48 kHz) |
| Speech enhancement | Denoisers, RNNoise-style pipelines |
| WebRTC / call audio | Full VoIP stack with codec + jitter buffer |
| Over-the-air recapture | Played through a speaker and re-recorded in a non-lab environment |
Background music mixing
A music track is mixed into the watermarked signal at a realistic SNR. This is the common "posted over a backing track" scenario.
Detection Result
Gaussian noise
White noise added at 20 dB SNR — simulates microphone hiss, noisy transmission, or deliberate masking attempts.
Detection Result
Cuts / splicing
A short segment is removed from the middle of the clip and the two halves are rejoined — simulates edits, ad breaks, or clip re-use.
Detection Result
Why a Low False-Positive Rate Matters
A detector is only as useful as its negative answer. If Vigil returned TRUE on content it never watermarked, the positive signal would be worthless.
Provenance decisions are asymmetric:
- False positive (claiming unrelated audio as yours) is a credibility problem — it can lead to wrongful attribution, mistaken takedown requests, or legal exposure.
- False negative (missing one of your own watermarked clips after a harsh attack) is a missed detection — recoverable, and usually still detectable on a different segment.
Vigil is tuned for an extremely low false-positive rate across a wide range of unrelated audio. When Vigil says TRUE, it means the watermark is genuinely present.