Mivia Sign writes a cryptographic watermark directly into the
DCT coefficients of your audio. Inaudible. Robust to re-encoding.
Byte-preserving on MP3. Timestamp-anchored to Bitcoin.
↓
Scroll to see how
+1.0
0
−1.0
stage01 / 07
scale1×
samples8,912,384
duration3:22.500
01 / 07
Your audio, full length.
A typical 3-minute stereo track at 44.1 kHz is 8.9 million
samples per channel. Mivia Sign treats it as a stream of
independent frames — 4096 samples each, 93 ms of audio at
the engine's 22,050 Hz working rate.
format
WAV · FLAC · AIFF · MP3
channels
mono or stereo
preserves
sample rate · bit depth · container
02 / 07
Zoom in. One frame.
Each frame carries the full watermark payload independently.
Every frame is a fresh chance to recover the signature — so
even aggressively cropped audio still verifies, as long as
one complete frame survives.
frame length
4,096 samples · 93 ms
payload
full 20-byte envelope, BCH-encoded
redundancy
every frame holds a copy
03 / 07
Move to the frequency domain.
The frame goes through a discrete cosine transform, splitting
it into 2,048 frequency coefficients. Human hearing is most
sensitive to mid-range frequencies — so the watermark skips
those and hides in the perceptually quiet bands above.
transform
DCT-II · orthonormal
coefficients
2,048 per frame
target band
above the masking threshold
04 / 07
A keyed subset of bins.
A blake3-derived master key picks which coefficients hold
the watermark — roughly 120 bins out of the 2,048. Without
the key, even a forensic listener can't find them; with
the key, extraction is O(1) per frame.
selection
blake3(platform_key) → bin indices
carriers per frame
~120
total bits per frame
160 (20 bytes × 8)
05 / 07
Embed one bit. Imperceptibly.
Each carrier coefficient is nudged to the nearest point on a
quantisation lattice with spacing δ = 0.002. That's about
−54 dBFS — below the noise floor of any practical listening
environment, and well below the threshold of audibility.
method
keyed QIM (quantisation index modulation)
step size
δ = 0.002
peak distortion
−54 dBFS (inaudible)
06 / 07
The envelope.
The 20 bytes embedded into every frame: a magic byte, a
format version, a 4-byte issuer fingerprint, an 8-byte
content hash pointing at the full manifest in the
content-addressed store, a random nonce, and a 16-bit CRC.
BCH-coded into 70 bytes before embedding so majority-vote
recovery tolerates up to 40 bit errors per frame.
byte 0
FB (magic)
byte 1
01 (version)
bytes 2–5
issuer fingerprint
bytes 6–13
contract hash (pointer)
bytes 14–17
nonce
bytes 18–19
CRC-16/CCITT
07 / 07
Your file, signed.
The output is byte-identical to your input except for the
imperceptible modifications inside the DCT coefficients. WAV
stays WAV at the same sample rate and bit depth. MP3 stays
MP3 at the same bitrate with the same frame count. Nothing
about the file's identity, structure, or listening experience
changes — except that it now carries your signature.
size delta
0 bytes (MP3) · unchanged (WAV)
perceptual delta
none
cryptographic delta
one 20-byte envelope per frame
Properties
−54 dBFS
Peak distortion
Inaudible on any consumer playback system. Below the
masking threshold of any non-silent program material.
20 bytes
Payload per frame
Magic, version, issuer, contract hash, nonce, CRC. Every
frame carries the complete signature.
0 bytes
MP3 size change
Compressed-domain scalefactor editing. The file is
byte-identical outside of the bits we intentionally flip.
40 bits
Error-correction headroom
BCH(1023, 463) ECC. Tolerates heavy re-encoding, noise
addition, and bit-level corruption before a frame fails.
BTC
Timestamp anchoring
OpenTimestamps proof registered at sign time. Independently
verifiable proof-of-existence backed by Bitcoin consensus.
v1
Frozen forever
The v1 algorithm bundle is locked via CI-gated constants.
Every signed file works forever regardless of how the
protocol evolves.
Sign your first file.
Free, no account. Works with WAV, FLAC, AIFF, and MP3.