Irodori-TTS Watermark Detector
Detects the SilentCipher watermark (payload IRDTS) that
irodori_tts/watermark.py embeds into generated speech.
Everything runs locally in your browser — inference executes in WebAssembly and no
audio ever leaves this machine.
How it works & how to read the numbers
Pipeline. The file is decoded with WebAudio at 48 kHz, downmixed to
mono, resampled to 44.1 kHz with a torchaudio-compatible sinc kernel, scaled to
SilentCipher's reference power, and transformed with a 4096-point STFT (hop 2048).
The first 1024 magnitude bands are fed to SilentCipher's message decoder
(dec_m_0, exported to ONNX) running on onnxruntime-web's WebAssembly
execution provider. It emits a distribution over 5 symbols for each frame.
Statistic. A watermarked signal repeats a fixed 21-symbol word
(20 payload symbols of two bits + a terminator) once per 21 frames. We score the known
IRDTS word at all 21 rotations; the 20 wrong rotations give an empirical
null distribution, so the test calibrates itself to the material. payload match
is the frame agreement at the best rotation, compared against that null; the
p-value is the corresponding binomial tail.
Watermark structure is payload independent: it measures how strongly the per-frame predictions repeat with period 21. High structure with a low payload match means a SilentCipher watermark is present but carries a different payload.
Limits. About 4 s of audio is needed before a verdict is issued;
longer is better. A payload that differs from IRDTS in only one byte still
correlates strongly, so always check the recovered payload and the symbol agreement
before treating a detection as proof of Irodori-TTS origin.