Save-my-song script

Invisible 16-bit watermark protection for your audio tracks, powered by Meta's AudioSeal.

Embed an inaudible watermark into any WAV file to prove ownership — or verify whether a file already carries one. One script, two modes, zero compromise on audio quality.

---

✨ Features

🎵 Embed mode	Inject a 16-bit watermark that's imperceptible to human hearing
🔍 Check mode	Detect and decode a watermark from any audio file
📧 Ownership Proof	Hash your email into the 16-bit payload to cryptographically tie the file to you
🎨 Pretty CLI	Colored output, progress feedback, timing, and file stats
⚙️ Flexible	Custom output paths, adjustable detection thresholds, quiet mode
🛡️ Robust	Input validation, graceful error handling, meaningful exit codes

📦 Requirements

• Python 3.12+
• PyTorch and torchaudio
• AudioSeal

pip install torch torchaudio audioseal

Note

Models are downloaded automatically from HuggingFace on first run and cached locally.

🚀 Usage

Embed a watermark with ownership proof

# Watermark a track and bind it to your email
python Save_my_song.py my_track.wav --email owner@example.com

# Specify a custom output path
python Save_my_song.py my_track.wav -o sealed_track.wav --email owner@example.com

Check for a watermark & verify ownership

# Detect watermark in a file
python Save_my_song.py --check sealed_track.wav

# Verify if a specific email matches the embedded ownership hash
python Save_my_song.py --check sealed_track.wav --verify-email owner@example.com

# Use a stricter detection threshold (0.0–1.0)
python Save_my_song.py --check sealed_track.wav -t 0.7

Example output

Embed mode:

   ╔═══════════════════════════════════════╗
   ║   🔐  AudioSeal  — Save_my_song.py    ║
   ║   Invisible watermark protection      ║
   ╚═══════════════════════════════════════╝

  ℹ Loading audio: my_track.wav
  ℹ Sample rate 44100 Hz  •  channels 2  •  duration 198.35 s
  ℹ Loading generator model …
  ✔ Generator ready
  ℹ Hashing ownership email → 16-bit payload …
  ✔ Ownership hash ready for: owner@example.com
  ℹ Generating watermark …
  ✔ Watermark generated in 3.21 s
  ℹ Saving → protected_my_track.wav
  ✔ Protected file saved (34.2 MB)

  ✔ Done! Your track is now protected. 🎶

Check mode:

  ℹ Loading audio: protected_my_track.wav
  ℹ Sample rate 44100 Hz  •  channels 2  •  duration 198.35 s
  ℹ Loading detector model …
  ✔ Detector ready
  ℹ Analysing watermark (threshold = 0.5) …

  ┌──────────────────────────────────────
  │  Detection probability : 0.9987
  │  Threshold             : 0.5
  │  Decoded 16-bit hash   : 1010110011001101
  │  Email verification    : MATCH ✔ (owner@example.com)
  │  Result                : WATERMARK DETECTED ✔
  └──────────────────────────────────────
  ℹ Analysis completed in 1.84 s

⚙️ CLI Reference

usage: Save_my_song.py [-h] [-o OUTPUT] [--check] [--email EMAIL] [--verify-email EMAIL] [-t THRESHOLD] [-q] input

Argument	Description
input	Path to the input WAV file
-o, --output	Output path for watermarked file (default: protected_<input>)
--check	Detection mode — check for a watermark instead of embedding
--email	Owner email to embed as a 16-bit ownership hash (embed mode)
--verify-email	Email to verify against the detected watermark hash (check mode)
-t, --threshold	Detection threshold, 0.0–1.0 (default: 0.5)
-q, --quiet	Suppress the banner and informational messages

📤 Exit Codes

Code	Meaning
0	Success — watermark embedded or detected
1	Error — missing file, invalid arguments, runtime failure
2	No watermark found (check mode only)

Tip

Exit code 2 makes it easy to script detection checks:

if python Save_my_song.py --check track.wav -q; then
    echo "Watermarked ✔"
else
    echo "Not watermarked"
fi

🧠 How It Works

AudioSeal uses a neural network to generate a frequency-domain watermark that sits below the threshold of human perception. The watermark survives common audio transformations (compression, re-encoding, minor edits) while remaining completely inaudible. Your ownership watermark can still be present if a file is converted to a mp3 and/or shortened (but the beginning of the song is intact).

┌────────────┐     ┌───────────┐      ┌─────────────────┐
│  Original  │ ──▶ │ Generator │ ──▶  │  Watermarked    │
│  audio.wav │     │  (embed)  │      │  protected.wav  │
└────────────┘     └───────────┘      └─────────────────┘
                                              │
                                              ▼
                                      ┌───────────┐
                                      │ Detector  │ ──▶  ✔ / ✖
                                      │  (check)  │      + decoded bits
                                      └───────────┘

• Generator model — audioseal_wm_16bits — embeds 16 bits of hidden information
• Detector model — audioseal_detector_16bits — recovers the probability + message

🔐 About the 16-bit Email Hash

AudioSeal's payload capacity is strictly 16 bits (65,536 combinations). An email address cannot fit directly. This script solves it by:

1. Hashing your email with SHA-256
2. Taking the first 16 bits as a deterministic fingerprint
3. Embedding that fingerprint into the audio
4. During verification, re-hashing the claimed email and comparing it to the decoded bits

This provides a lightweight, reproducible proof-of-ownership mechanism suitable for creator tracking and attribution. For enterprise-grade cryptographic guarantees, consider pairing this with external metadata registries.

2026 [ ivan deus ]