Audio Watermarking Lexicon
How Aureal embeds, survives and reports audio identifiers — with the limits stated. All 30 planned topics.
-
What is audio watermarking?Basics
Audio watermarking embeds a keyed ID in the audio itself. Aureal uses spread spectrum; a match is a provenance signal, not proof of ownership.
-
Watermark vs. metadataBasics
Metadata is a removable label; a watermark lives in the audio samples. Learn what survives tag stripping, re-encoding and platform processing.
-
Watermark vs. fingerprintingBasics
A fingerprint derives an ID from the audio; a watermark inserts one. Aureal inserts a 32-bit ID. Neither approach is collusion-proof.
-
Watermark vs. DRMBasics
DRM controls access; a watermark records an identifier. Aureal does not prevent copying or playback. Use each for its own job.
-
Payload ID and keyBasics
Aureal carries a 32-bit ID plus CRC. Detection needs the same key. Re-marking with the same key and a new ID can destroy the original.
-
MP3 128 kbpsRobustness
MP3 128 kbps passed on Aureal's local fixture matrix. That is a fixture result, not a guarantee for every encoder, source or platform.
-
MP3 320 kbpsRobustness
MP3 320 kbps passed on Aureal's local fixture matrix. Higher bitrate helps, but recovery still depends on source, encoder and processing.
-
AAC transcodingRobustness
AAC 128 kbps passed on Aureal's local fixture matrix. AAC encoders vary; validate the exact bitrate and encoder you use.
-
Opus transcodingRobustness
Opus is in Aureal's tested codec family. There is no separate per-bitrate guarantee; test the Opus setting and encoder you actually use.
-
ResamplingRobustness
Aureal retries detection at 44.1 kHz and 48 kHz after the native rate. Other rates need their own test; resampling is not a universal guarantee.
-
Cropping and trimmingRobustness
Aureal locks onto a keyed chirp preamble, so unaligned crops can decode. On the named 12 s fixture, 1.8 s excerpts recovered at half the offsets.
-
Speed and pitch driftRobustness
Aureal retries detection at ±0.2%, ±0.5% and ±1.0% speed drift. That covers small linear changes, not tape wow, varispeed edits or large tempo shifts.
-
Stereo to mono downmixRobustness
Aureal embeds in the mid component, so an in-phase mono downmix adds about 3 dB. Out-of-phase content and widening can still reduce or cancel it.
-
Clipping and limitingRobustness
Hard clipping at +12 dB passed a local fixture, as did one heavy limiter. Both are nonlinear; test the final master at its release loudness.
-
Background noiseRobustness
Noise fixtures passed down to −30 dBFS. In-band SNR decides, broadband noise is a weak host, and near-silence is muted rather than marked.
-
8 kHz telephonyRobustness
At 8 kHz the standard carriers do not fit under Nyquist, so Aureal uses a ~2–4 kHz fallback. It is a known weak case; test before relying on it.
-
Short excerptsRobustness
A frame is about one second. Sub-1 s excerpts never decoded; at 1.0–1.8 s recovery was partial and depended on the start offset in the named fixture.
-
HMAC seal vs. signatureForensics
No. Aureal seals reports with HMAC-SHA256, a shared-secret MAC. It detects tampering and shows key possession, not identity, authorship or time.
-
Verify a forensic reportForensics
Run verify-report with the same shared secret. Exit 0 means the body is unchanged; exit 1 means modified or wrong key. It does not prove authorship.
-
SHA-256 and SHA-512Forensics
They fingerprint the scanned file's bytes. A match shows the file is unchanged; it does not establish authorship, ownership or creation time.
-
A match is not authorshipForensics
No. A match shows the audio carries a CRC-valid ID tied to a key and a distribution record. It does not establish who made, owns or leaked it.
-
BER and confidenceForensics
BER is the share of the 48 codeword bits that differ from the expected ID; confidence is a normalized z-score. Blind-mode BER is always 0 by design.
-
Blind vs. verify modeForensics
Blind detection recovers any CRC-valid ID under the key; verify checks an expected ID and reports a real BER. Use verify when you have the manifest.
-
Batch distributionWorkflow
Embed one ID per recipient and keep a private recipients.json with the key and SHA-256 per file. Never publish it; averaging still defeats recovery.
-
Leak auditWorkflow
Run blind detection on the leaked file, then look up the ID in your private manifest. A hit names a distribution record, not the person who leaked it.
-
Key managementWorkflow
The key is an obfuscation-grade PRNG seed, not an encryption key. Lose it and detection fails; publish it and anyone can reproduce your IDs.
-
EU AI Act Article 50Compliance
Article 50 requires providers to mark synthetic audio machine-readably from 2 August 2026. A watermark is one measure; it is not a compliance certificate.
-
C2PA and watermarkingCompliance
C2PA Content Credentials are signed provenance manifests attached to a file; a watermark lives in the samples. They differ, and can be combined.
-
AI audio and deepfakesCompliance
No. Aureal detects its own embedded watermark, not synthetic speech. A missing mark does not mean human; a present mark does not mean AI.
-
AI voice providersUse case
Yes, as an integration: embed a per-generation ID after synthesis, keep the mapping private, and test the vocoder, codec and 8 kHz paths.
Scope and limits
- Detection depends on source audio, codec and processing.
- Sub-1 s excerpts never decode; ~1.8 s crops recover about half the time in local fixtures.
- 8 kHz and broadband-noise hosts are weak.
- Collusion/averaging defeats recovery; no collusion-resistant fingerprinting is implemented.
- A match is not authorship, ownership, creation time or chain of custody.
- The product detects its own watermark; it is not AI detection or deepfake detection.
Have a file in hand? Test the exact codec, bitrate and processing chain you use.
Open the Studio → Read the documentation