Methodology

Analysis Methods

Transparency in methodology is non-negotiable. Every technique we use is documented here so our work can be scrutinized and replicated.

truth@evades:~$ forensic-analysis

00 // Forensic Acoustic Pipeline

Our analyses follow a standardized 10-phase methodology designed for reproducibility. Each phase produces documented outputs that feed the next, creating a verifiable chain from raw evidence to final conclusions.

Intake — Evidence cataloging, hash verification, chain of custody

Characterization — Format analysis, codec inspection, recording parameters

Environment — Temperature, humidity, barometric pressure, venue geometry

Classification — Event type ID, shot counting, source separation

Geolocation — Receiver positioning via video/photo correlation, Google Earth Pro

Synchronization — GCC-PHAT cross-correlation sync (±21μs demonstrated precision)

TOA — Time-of-arrival onset picking (manual + automated, cross-validated)

TDOA Multilateration — Hyperbolic solver with iterative least-squares refinement

Uncertainty — Monte Carlo (10,000+ iterations), jackknife stability, sensitivity analysis

Reporting — Full methodology documentation, reproducibility as the standard

Core Methods // Detail

Time Difference of Arrival (TDOA)

When a sound event is captured by multiple recording devices at known positions, the time difference between arrivals constrains the source location to a hyperbolic curve. With three or more receivers, multilateration pinpoints the source position through hyperbolic intersection geometry.

▸ Synchronization: GCC-PHAT (Generalized Cross-Correlation with Phase Transform) for sub-sample timing alignment — demonstrated ±21μs precision in FA-2026-001

▸ Speed of sound: Cramer (1993) model incorporating temperature, relative humidity, barometric pressure, and CO₂ concentration for environment-corrected propagation velocity

▸ Solver: Iterative least-squares minimization on TDOA residuals with hyperbolic intersection geometry. Residual analysis flags degenerate receiver configurations

▸ Uncertainty: Monte Carlo simulation (10,000+ iterations) perturbing onset times, receiver positions, and environmental parameters. Output: 95% confidence ellipses and jackknife leave-one-out stability tests

Applications: shooter localization, shot count verification, source separation

Tools: custom Python TDOA solver, GCC-PHAT sync engine, Google Earth Pro (KML overlays), Praat, Audacity

Spectral Analysis

Fourier transform-based decomposition of audio signals into frequency components over time. Spectrograms reveal patterns invisible in the time domain — harmonics, formants, transient events, and compression artifacts all have distinct spectral signatures.

▸ Mel-frequency analysis: MFCC extraction for acoustic event characterization, matching perceptual frequency scales used in the gunshot classifier's feature pipeline

▸ Wavelet decomposition: Multi-resolution time-frequency analysis for transient event isolation — superior to fixed-window STFT for impulsive gunshot signatures

▸ Classifier interpretability: SHAP (SHapley Additive exPlanations) feature importance analysis showing which spectral bands drive classification decisions — essential for forensic transparency

Applications: sound classification, editing detection, environmental characterization, caliber discrimination

Tools: Praat, Audacity, SoX, custom Python (scipy, librosa, SHAP, scikit-learn)

Impulse Response & Echo Analysis

Impulsive sounds (gunshots, claps, etc.) reveal the acoustic properties of the recording environment through their reflections. By analyzing the timing and amplitude of echoes, we can distinguish direct sounds from reflections — critical for accurate shot counting and source characterization.

▸ Crack-to-bang separation: Temporal isolation of supersonic crack (shockwave from projectile) from muzzle blast (propellant gas expansion). The time gap constrains shooter-to-microphone distance

▸ Mach cone geometry: For supersonic rounds, the shockwave propagates as a cone with half-angle θ = arcsin(c/v). The arrival angle relative to the muzzle blast constrains the bullet trajectory vector

▸ N-wave detection: Identification and characterization of the bipolar pressure signature from supersonic projectiles — rise time, duration, and amplitude ratio discriminate caliber classes

Applications: shot counting, environment modeling, source/reflection separation, supersonic confirmation

Tools: Praat, custom convolution analysis, RT60 measurement, N-wave detector

Metadata & Codec Forensics

Audio file metadata (container format, codec parameters, encoding history) provides provenance information. Compression artifacts, re-encoding signatures, and container inconsistencies can reveal whether a file has been modified, re-encoded, or assembled from multiple sources.

▸ Re-encoding detection: x264 signature identification in video-sourced audio — detects screen recording, social media re-compression, and editing software artifacts

▸ Container/codec mismatch: Analysis of container format (MP4/MOV/MKV) vs. internal codec parameters to identify re-muxing, transcoding, or file manipulation

▸ Atom inspection: EXIF and QuickTime atom-level examination for creation timestamps, device signatures, GPS data, and encoding software identification

Applications: authenticity verification, chain-of-custody analysis, re-encoding detection, provenance tracking

Tools: MediaInfo, FFprobe, ExifTool, hex editors, custom Python atom parsers