How SonicID works
Every voice is a fingerprint written in mathematics. SonicID decodes it.
Capturing the signal
When you speak, your vocal cords vibrate at a fundamental frequency — the pitch your ear perceives. SonicID captures this raw waveform at up to 48 kHz, giving us 48,000 amplitude samples per second to work with. Every nuance of your voice — breathiness, resonance, texture — lives inside that stream of numbers.
The Fast Fourier Transform
A raw waveform tells us amplitude over time. To understand timbre we need amplitude over frequency — and that's exactly what the FFT delivers. SonicID runs a Cooley-Tukey radix-2 FFT on up to 32,768-point windows, converting each frame from the time domain into a spectrum of frequency bins. A Hann window is applied before each transform to suppress spectral leakage at frame boundaries, keeping the bins clean and accurate.
Pitch tracking via autocorrelation
Finding the fundamental frequency (F0) from a noisy spectrum is harder than it sounds — harmonics, noise, and octave errors all conspire against naive peak-picking. SonicID uses time-domain autocorrelation: the signal is correlated with a delayed copy of itself, and the first strong peak in the lag domain reveals the true period. This approach is robust to missing fundamentals and handles voiced speech reliably across the full human range of 80–1100 Hz.
Harmonic extraction
Once F0 is known, harmonics are integer multiples of it — 2×F0, 3×F0, and so on. SonicID extracts up to 16 partials, measuring the amplitude of each harmonic relative to the fundamental. This harmonic series is the tonal DNA of your voice: a rich second harmonic gives warmth, strong odd harmonics add edge, and the roll-off rate determines brightness. The pattern is unique to you.
Formant detection
Formants are resonant peaks in the vocal tract's frequency response — shaped by the position of your tongue, lips, and jaw. F1 (300–1000 Hz) encodes vowel height; F2 (700–2500 Hz) encodes vowel frontness; F3–F5 add the personal coloring that makes your voice recognizable even when whispering. SonicID detects F1 through F5 by finding spectral envelope peaks above the harmonic series, giving a five-dimensional map of your vocal tract geometry.
Energy band analysis
The spectrum is divided into six perceptually meaningful bands: sub-bass (20–80 Hz), bass (80–250 Hz), low-mid (250–500 Hz), mid (500–2000 Hz), high-mid (2–6 kHz), and presence (6–20 kHz). The relative energy in each band describes the overall character of your voice — whether it sits in the chest, the mask, or the head — and how it will translate across different playback systems.
The voice fingerprint
All extracted features — fundamental, harmonics, formants, spectral centroid, rolloff, flatness, and band energies — are combined into a compact feature vector and hashed into a 64-character hex fingerprint. This hash is deterministic: the same voice under the same conditions always produces the same hash, making it useful for identity comparison without storing raw audio.
The math at a glance
Discrete Fourier Transform — the foundation of every spectrum SonicID computes. N is the window size (up to 32,768 points); k is the frequency bin index; x[n] is the windowed time-domain sample.
Voice classification
Fundamental frequency ranges used to classify voice type
Ready to decode your voice?
Upload a sample or record directly in your browser — results in seconds.
Analyze my voice