The same source signal reaches each microphone with a different delay that depends on the array geometry and the source direction. TDOA methods estimate these inter-microphone delays, typically by maximizing GCC-PHAT (cross-correlation with phase-spectrum whitening, robust to reverberation), then derive the angle/position from the array geometry. SRP-PHAT scans a grid of candidate directions, summing the "steered" signal power and picking the maximum. MUSIC builds the covariance matrix of the signals, splits the space into signal and noise subspaces via eigendecomposition, and the peaks of its pseudo-spectrum (orthogonality of steering vectors to the noise subspace) reveal the directions. Deep learning approaches feed the phase and magnitude of all channel spectrograms into a CNN/CRNN that learns to map them directly to azimuth and elevation (regression or classification over an angular grid).
A single microphone records only the amplitude of sound over time, not its direction. SSL solves the problem of recovering spatial information (where the sound comes from) from multiple audio streams, which is essential for a machine to orient itself in an acoustic scene, separate a speaker from background noise and direct attention toward them.
A set of multiple microphones arranged in a known geometry (linear, circular, spherical). The time/phase differences between channels are the raw material for all SSL methods. More microphones increase resolution and robustness.
Generalized cross-correlation with phase transform. It whitens the spectrum (normalizes magnitude, keeps phase) so that a sharp correlation peak corresponds to the delay between a microphone pair even under reverberation. The basis of TDOA methods.
Official
Steered response power with phase transform. It scans a grid of candidate directions/positions, for each one sums the phase-weighted correlations of all microphone pairs, and selects the global maximum. Robust to reverberation but computationally costly.
Official
Multiple Signal Classification. It builds a covariance matrix and splits it into signal and noise subspaces via eigendecomposition; the pseudo-spectrum peaks (orthogonality of steering vectors to the noise subspace) indicate the directions. High resolution but requires knowing the number of sources.
Official
A CNN/CRNN network that takes the magnitude and phase of all channel spectrograms and learns to map them directly to azimuth/elevation. Handles reverberation and multiple sources better than classical methods but requires training data.
Official
Wall reflections create phantom sources and blur correlation peaks, reducing accuracy. PHAT mitigates but does not fully eliminate this.
Overlapping sources make peak assignment hard; MUSIC needs their count, and TDOA methods may confuse sources.
Too-large microphone spacing causes spatial aliasing; linear arrays cannot distinguish front from back.
Knapp and Carter formalize the GCC family of estimators with phase weighting (PHAT), the foundation of TDOA methods.
R. O. Schmidt publishes MUSIC in IEEE Trans. Antennas and Propagation, enabling high-resolution DOA estimation for multiple sources.
J. DiBiase (PhD dissertation, Brown University) introduces SRP-PHAT for talker localization in reverberant rooms.
CNN/CRNN networks (e.g. DOAnet, Adavanne et al.) estimate DOA directly from multichannel spectrograms, improving robustness to reverberation.
Adavanne et al. combine acoustic event detection with DOA estimation in a single CRNN, launching the SELD task (and the DCASE challenge).
Time complexity: O(G · M²). Space complexity: O(M² + G).
In steered-response methods the main cost is evaluating power over a dense candidate grid; in MUSIC it is the eigendecomposition; in DNN approaches it is network inference.
More microphones increase angular resolution and the number of separable sources but raise cost and complexity.
The microphone layout (linear, circular, spherical) determines azimuth/elevation coverage and possible ambiguities (front-back).
Density of the candidate-direction grid in steered-response methods — a trade-off between accuracy and compute cost.
Subspace methods (MUSIC) require knowing/estimating the number of active sources; an error degrades the result.
Classical methods process all channels and grid points; there is no conditional routing.
Computations for individual direction-grid points and microphone pairs are independent, so SRP-PHAT parallelizes well. DNN estimation benefits from GPU parallelism.
Classical methods (GCC-PHAT, SRP-PHAT, MUSIC) rely on FFT and matrix algebra well vectorized on CPU; they run in real time on a robot embedded processor.
Deep learning approaches (CNN/CRNN) for DOA estimation benefit from GPU acceleration for both training and inference.
