Check the mark yourself

chatFlow marks AI voices with AudioSeal, an open method by Meta (MIT license). This is how you check a file on your own computer, without our site.

1. Install the tool (Python 3):

pip install "audioseal>=0.2.0" soundfile

2. Convert the recording to mono at 16 kHz:

ffmpeg -i aufnahme.mp3 -ac 1 -ar 16000 pruefen.wav

3. Check:

import sys, soundfile as sf, torch
from audioseal import AudioSeal

CHATFLOW = "1011001011100101"
wav, sr = sf.read(sys.argv[1], dtype="float32")
x = torch.from_numpy(wav)[None, None, :]
detector = AudioSeal.load_detector("audioseal_detector_16bits")

for d in range(0, 320, 4):
    anteil, bits = detector.detect_watermark(x[..., d:], sample_rate=sr)
    nachricht = "".join(str(b) for b in (bits[0] > 0.5).int().tolist())
    if float(anteil) > 0.5 and nachricht == CHATFLOW:
        print(f"chatFlow ({float(anteil):.2f}, {d})")
        break
else:
    print("-")

python pruefen.py pruefen.wav

Our mark always carries the same 16-bit message: 1011001011100101. It counts as found when more than half of the frames are above 0.5 and all 16 bits match.

Why the loop: the network works in steps of 320 samples. A freely cut piece usually sits beside this grid; without shifting it is often not detected.

← Back to the check page