Author Note: AI wrote this. Don't have the time at this moment for Readmes. Soon.
Professional audio forensics analysis with calibrated thresholds for real-world mastered audio.
A command-line audio forensics tool that analyzes audio files to determine their authenticity, quality, and technical characteristics. It combines multiple audio analysis engines (ffmpeg, SoX, mediainfo) with a custom numpy-based spectral analysis engine to provide comprehensive forensic insights.
Unlike generic audio analysis tools, Audio Forensic was built with real-world mastered audio in mind — not synthetic test signals.
Most audio forensics tools are calibrated against:
- Clean sine waves
- White/pink noise
- Synthetic test signals
This causes massive false positives when analyzing actual music:
| Metric | Generic Tool Says | Reality |
|---|---|---|
| DR5 | "BAD" (red) | Normal for modern pop/rock |
| Crest Factor 3dB | "BAD" (red) | Typical for compressed audio |
| SoX Entropy 0.09 | "Suspicious" (red) | Normal for tonal music |
| HF cutoff 19.8kHz | "Lossy!" (red) | Standard CD mastering |
| Peak at 0.999 | "Clipping!" (red) | Normal loudness normalization |
af2 fixes all of this.
- Technical Metadata: Bit depth, sample rate, channels, encoding
- Loudness Profiling: LUFS, DR, crest factor, EBU R128
- Spectral Analysis: FFT-based authenticity detection
- Dynamic Range: True peak, RMS, noise floor
- Phase Correlation: Stereo width and compatibility
- ReplayGain Audit: Tag verification
- Calibrated for commercially mastered audio
- Context-aware color coding
- Human-readable interpretations for every metric
- No false positives from normal mastering practices
- Clean ANSI terminal output
- JSON export for automation
- Batch mode for album analysis
- Spectrogram generation
# Core dependencies (must be in PATH)
ffmpeg # Audio decoding and analysis
sox # Statistical analysis
mediainfo # Metadata extraction
# Python dependencies
numpy # Spectral analysisbrew install ffmpeg sox mediainfo
pip3 install numpysudo apt install ffmpeg sox mediainfo
pip3 install numpyDownload and install:
Then:
pip install numpypython af2.py "path/to/audio.flac"python af2.py *.flacpython af2.py "track.flac" --jsonpython af2.py "track.flac" --fastpython af2.py "track.flac" --info── DYNAMIC RANGE & LOUDNESS ────────────────────────────────
DR Score (EBU) DR5 — normal for modern mastered audio
Crest Factor 2.94 dB — compressed (modern standard)
SoX Entropy 0.09 — very low (highly tonal/structured)
── AUTHENTICITY & FORENSICS ───────────────────────────────
████░░░░░░ ✓ Strong evidence of authentic lossless source
Score: Lossy 2 − Natural 6 = Net 0
Lossy indicators
· None detected
Natural indicators
· Gradual spectral rolloff (natural EQ / mastering)
· High cutoff variance (organic/analog source)
· Natural HF noise above cutoff
· Healthy stereo image
· High spectral complexity
Context notes
· Modern mastering often uses gentle HF limiting at 19-20 kHz
| Metric | What it means | Normal range |
|---|---|---|
| DR Score | Dynamic Range (higher = more dynamic) | DR5-12 for modern music |
| Crest Factor | Peak-to-RMS ratio | 3-15 dB depending on genre |
| SoX Entropy | Time-domain signal randomness | 0.0-0.3 for music |
| Metric | What it means | Suspicious if |
|---|---|---|
| HF Cutoff | Highest frequency with content | <18 kHz |
| Cliff Sharpness | How abrupt the rolloff is | >15 dB/bin |
| HF Energy Ratio | Energy above 15kHz | <0.005 |
| Noise Floor | Content above cutoff | <-70 dB (silent void) |
| Banding Score | Quantization artifacts | >0.92 with low cutoff |
| Label | Meaning |
|---|---|
| GENUINE | Strong evidence of authentic lossless |
| LIKELY_GENUINE | Consistent with lossless source |
| CAUTION | Minor quirks, likely legitimate |
| SUSPICIOUS | Spectral anomalies detected |
af2 is a complete rewrite with calibrated thresholds for real audio:
| Feature | af.py | af2.py |
|---|---|---|
| DR5 coloring | RED (wrong) | WHITE (correct) |
| Crest Factor coloring | RED (wrong) | WHITE (correct) |
| SoX Entropy interpretation | "Suspicious" | "Typical music" |
| Peak at 0.999 coloring | RED (wrong) | YELLOW (correct) |
| HF cutoff threshold | 97% Nyquist | 85% Nyquist |
| HF energy threshold | <0.02 suspicious | <0.005 suspicious |
| Metric explanations | "low=..., high=..." | "moderate variation (natural)" |
| Verdict accuracy | Many false positives | Calibrated for mastered audio |
Track 1.flac DR5 -14.8 LUFS ✓ GENUINE
Track 2.flac DR6 -14.2 LUFS ✓ GENUINE
Track 3.flac DR4 -15.1 LUFS ✓ GENUINE
af2 correctly identifies all tracks as genuine, while af.py flagged them as "WARNING" due to incorrect threshold calibration.
Source: FLAC 16/44.1 → MP3 128kbps → FLAC 16/44.1
af.py: WARNING (Net 4/17)
af2.py: SUSPICIOUS (Net 7) ✓ Correct
af2 correctly identifies the lossy source using calibrated HF and banding thresholds.
- DR5 is not bad — it's normal for music released after 1990
- Peaks at 0.999 are not clipping — they're normal mastering
- HF rolloff at 20kHz is not suspicious — it's standard CD mastering
- Low entropy is not transcode evidence — it's normal for tonal music
- True digital overs (>1.0 sample values)
- Silent voids above cutoff (not purple noise)
- Hard spectral cliffs (>15 dB/bin)
- Extremely stable cutoffs with very low HF
- Joint stereo artifacts in side channel
Issues and pull requests welcome! If you find a file that af2 misclassifies, please open an issue with:
- The file's characteristics (genre, year, label)
- The af2 output
- Any known history of the file
MIT License — do whatever you want with it.
- SoX team for the original stat tool
- ffmpeg team for audio processing
- MediaArea for metadata extraction
- The audio forensics community for threshold research
Made with Python, numpy, and too much listening to compressed music.