Acoustic Firearm Classification
A multi-iteration classification system trained on the CADRE ballistic audio dataset, modeled on research by Robert C. Maher. Classical machine learning on physics-based features outperformed a deep neural network approach — and the reasons why are as instructive as the results.
Foundation // Robert C. Maher
This work is modeled on the research of Robert C. Maher (Montana State University), whose publications on gunshot acoustics established the scientific basis for acoustic firearm analysis. Key findings from his research that shaped this classifier:
- ▸ Gunshot waveforms consist of separable components — muzzle blast, supersonic shockwave (N-wave), and reflections — each carrying discriminative information
- ▸ Shot-to-shot variability is significant even under controlled conditions (~10% peak pressure variation for rifles, higher for handguns)
- ▸ Consumer recording devices introduce 5–10ms time smearing on gunshot onsets (3+ meters of error at the speed of sound)
- ▸ "Conventional audio recordings have not been shown to be reliable for identifying particular firearms" — a limitation this project directly confronts
The Problem
Why This Matters
In forensic audio analysis, identifying the firearm used in an incident from audio recordings alone is a problem with real investigative value. Different calibers produce distinct acoustic signatures — muzzle blast characteristics, supersonic crack profiles, spectral energy distribution — but these differences are subtle and environment-dependent.
Why It's Hard
- ✗ Recording conditions vary wildly (distance, environment, device)
- ✗ Echoes and reflections contaminate the signal
- ✗ Different firearms can share the same caliber
- ✗ Consumer device microphones clip, compress, and apply AGC
- ✗ Domain shift: lab recordings don't sound like field recordings
Training Data // CADRE Dataset
Trained on the CADRE (Community Accessible Digital Resource for Forensics) ballistic audio dataset, combining two collection efforts: a multi-firearm, multi-orientation edge-collected dataset from the Center for Advanced Data Analytics & Systems (CADAS) and the Air Force Research Laboratory (AFRL) (Kabealo et al., Data in Brief), plus CADRE's own iPhone-recorded collection.
The CADAS/AFRL recordings were captured with multichannel microphone arrays (up to 7 channels) across multiple orientations and firing styles, then cropped to 2-second clips containing at least one gunshot. Metadata includes firearm, caliber, GPS, recording device, microphone, and per-shot timestamps.
Feature Engineering
Feature Distribution
Gunshot-Specific Features
Beyond standard audio features, the system extracts physics-based signatures derived from Maher's research on ballistic acoustics:
- ▸ Muzzle Blast Energy — Low-frequency (<500Hz) content from propellant gas expansion
- ▸ N-Wave Detection — Bipolar shockwave from supersonic projectiles. Rise time, duration, amplitude ratio
- ▸ Impulse Kurtosis — Statistical measure of signal impulsiveness (high for gunshots vs. other transients)
- ▸ Crack-to-Blast Ratio — High-frequency crack energy vs. low-frequency blast. Caliber-dependent
- ▸ Attack/Decay Profile — Rise time to peak and decay to -20dB. Differs by barrel length and caliber
SUPERSONIC vs SUBSONIC DISCRIMINATION
Generates an N-wave — a characteristic bipolar shockwave from the supersonic projectile's Mach cone. High-frequency content (>2kHz), rapid sign changes, detectable via high-freq kurtosis.
Muzzle blast only — no crack. Smooth low-frequency energy, longer decay, distinct spectral envelope. Absence of N-wave is itself a classification signal.
Two Approaches // ML vs Neural Network
The classifier was developed along two parallel tracks: classical machine learning on handcrafted features, and deep transfer learning via a pretrained neural network. The results reveal something important about this problem domain.
CLASSICAL ML
Physics-based feature engineering derived from Maher's research. 200+ handcrafted features targeting known ballistic acoustic phenomena — muzzle blast energy, N-wave characteristics, spectral band ratios, wavelet decomposition. The features encode domain knowledge about how gunshots actually work.
NEURAL NETWORK
Google's YAMNet — a deep neural network pretrained on AudioSet (10M+ hours of general audio) for sound event classification. Embeddings (3,072 dimensions) extracted and fused with acoustic features for a 4,116-dimension feature vector. YAMNet understands sound in general, but the subtle ballistic physics that distinguish calibers aren't in AudioSet's training distribution.
WHY CLASSICAL ML OUTPERFORMED THE NEURAL NETWORK
- 01 Domain knowledge matters more than data volume here. The handcrafted features encode acoustic physics (muzzle blast frequency content, N-wave timing, crack-to-blast ratios) that directly correspond to what makes calibers physically different. YAMNet's pretrained knowledge of "what a gunshot sounds like" is too general for caliber-level discrimination.
- 02 YAMNet memorized the dataset. 100% training accuracy with 55.6% validation = the network learned to recognize individual CADRE recordings, not generalizable ballistic signatures. With ~5,000 samples across 22 classes, there isn't enough data to fine-tune a deep network without massive overfitting.
- 03 Feature engineering is interpretable. With classical ML, you can inspect which features drive predictions (via SHAP). When the model says ".308" you can verify it's looking at the right spectral bands. A neural network's reasoning is opaque — a problem for forensic work where you need to explain your conclusions.
Evolution // Model Iterations
Each iteration addressed specific limitations discovered in the previous version. Click to expand.
Key insight: classifying by caliber instead of specific firearm dramatically improves accuracy. Calibrated ensemble with isotonic probability calibration for forensically reliable confidence scores. Highest cross-validation accuracy of any iteration.
- ▸Caliber-level classification (31 firearms → 14 calibers)
- ▸Ballistic-focused feature extraction (80+ features)
- ▸Isotonic probability calibration for honest confidence scores
- ▸Device confound removal (strips iPhone/array labels)
- ▸Ensemble voting with confidence assessment
The Hard Truth // Domain Shift
The best cross-validation accuracy on CADRE data is 83.5%. But when the classifier processes real-world forensic recordings (UVU evidence audio), confidence drops dramatically. This is the domain shift problem — the core challenge Maher's research predicted.
- Quasi-anechoic, controlled environment
- Multi-channel professional array
- Known firearm, distance, position
- Clean WAV, no compression
- Minimal background noise
- Reverberant, uncontrolled environment
- Single phone mic with AGC and clipping
- Unknown distance, angle, obstructions
- Compressed, re-encoded, shared via social media
- Wind, traffic, crowd noise, overlapping events
While caliber-level classification confidence was low, the classifier's N-wave detection returned a definitive result: all 8 successfully processed UVU recordings contain supersonic shockwave signatures (1 of 9 files errored during processing). This is a physics-based detection — not a statistical prediction — and it immediately constrains the problem space. The rounds fired were supersonic, ruling out subsonic calibers (.45 ACP, standard .38 Special, 12 gauge slugs). This is the classifier operating at its most reliable: binary acoustic physics, not probabilistic classification.
The caliber classification confidence being low isn't a failure of the model — it's the model honestly reporting that it's outside its training distribution. The isotonic calibration ensures the confidence scores mean something: when it says "low confidence," it means it. That honesty is more valuable than false certainty. Meanwhile, the physics-based detections (N-wave, muzzle blast characterization) still deliver actionable findings regardless of domain shift.
Synthetic Augmentation
The GunShotSynthesizer generates realistic acoustic variations of training samples, expanding the dataset 5× while simulating real-world recording conditions that the clean CADRE recordings don't capture. This is one approach to narrowing the domain gap.
- Anechoic (reference)
- Outdoor open / urban
- Indoor small / large
- Parking garage
- Forest / Canyon
- Inverse square law attenuation
- Atmospheric absorption
- Mach cone delay calculation
- Barrel length effects
- Suppression modeling
- AWGN (0–100 dB SNR)
- Environmental noise mixing
- Mic response simulation
- Directional patterns
- SNR-controlled blending
Stack
This classifier is a research tool — not a definitive forensic instrument. Classical ML with physics-based features outperformed deep learning here because domain knowledge matters more than model complexity when training data is limited and the problem is grounded in known physics. All classifications include calibrated confidence scores, and results should be interpreted alongside traditional acoustic analysis methods.