Source-linked AI summary
Conformal Prediction-Aware DOA Tracking
Maya Veisman Barness, Sharon Gannot, Bracha Laufer Goldshtein
TL;DR
Deep DOA estimators can produce uncalibrated probabilities, particularly in noise, while temporal tracking and conformal prediction have been studied separately. The paper combines neural posterior estimation, conformal calibration, and confidence-weighted temporal smoothing, reporting improved robustness and stability, particularly under burst noise.
Problem
Deep neural DOA estimators may produce uncalibrated, overly confident probabilities in noise, while prior work has considered temporal tracking and conformal uncertainty quantification in localization.
Source: §1 · p. 1Method
The framework feeds microphone iRTFs to a neural DOA estimator, calibrates its frame-wise predictions with conformal prediction, and blends its posterior with a temporally propagated belief using conformal confidence.
Source: §2.1 · p. 1, §2.2 · p. 1, §2.3 · p. 2Results
Takeaways & Limitations
In the reported experiments, conformal-aware smoothing suppresses spurious DOA estimates from directional bursts while preserving the moving source's underlying trajectory.
Source: §3 · p. 2Because calibration uses static scenes and testing uses a moving-speaker scenario, the exchangeability assumption holds only approximately on test data.
Source: §2.2 · p. 1
Abstract
from arXiv · showhide
We introduce an uncertainty-aware DOA-tracking framework combining deep posterior estimation, conformal-prediction calibration, and temporal smoothing. The approach enhances stability against reverberation and burst-like directional noise, yielding more reliable trajectories and greater robustness than conventional DOA-tracking methods.
1. INTRODUCTION
DOA estimators can be highly accurate, but their probabilities may be uncalibrated and overconfident in noise. This work integrates conformal prediction into DOA tracking to use confidence in adaptive temporal smoothing.
1. INTRODUCTION
- Deep neural DOA estimators have demonstrated high accuracy, but may produce uncalibrated, overly confident probabilities, especially in noise.Source: §1 · p. 1
- Prior work has studied temporal continuity in tracking and conformal prediction for uncertainty quantification in localization.Source: §1 · p. 1
- The framework combines a neural estimator, conformal prediction, and a DOA time-smoother whose balance between neural posteriors and temporal continuity is guided by conformal-set confidence.The authors report improved accuracy and robustness under reverberation and noise bursts.Source: §1 · p. 1
2. METHODOLOGY
The method estimates DOA from microphone-signal iRTFs, calibrates frame-wise neural predictions with conformal prediction, then smooths them using temporal continuity weighted by conformal confidence.
2.1. Signal Model and iRTF Representation
- Complex iRTFs, represented by their real and imaginary parts for microphones other than the reference, form the DNN input for instantaneous DOA estimation.The iRTF is defined as a ratio between acoustic transfer functions and can be estimated from microphone signals.Source: §2.1 · p. 1
2.2. Neural Network Architecture and Conformal Calibration
- A 2D U-Net estimates frame-wise softmax probabilities over 32 candidate angles, with angles within 10° of endfire excluded.The network uses four downsampling and four upsampling blocks with skip connections; the maximum-probability angle is the instantaneous estimate.Source: §2.2 · p. 1
- Conformal prediction calibrates the neural estimator using a separate calibration set, with the conformal set expanding in uncertain or noisy conditions and shrinking when the model is confident.Training and calibration use static scenes, while evaluation uses a moving-speaker scenario, so exchangeability holds only approximately for test data.Source: §2.2 · pp. 1–2
2.3. DOA Tracking with CP-Aware Weighting
- A row-stochastic Gaussian transition matrix propagates the previous DOA belief to a one-step prior, with σc controlling the spread of allowable transitions.Source: §2.3 · p. 2
- The smoother blends the propagated prior with the neural posterior into a belief over DOA classes, then selects the class with the highest probability.Source: §2.3 · p. 2
- The smoothness score uses three descriptors of the conformal set, with small contiguous sets increasing the weight and large or irregular sets decreasing it.The descriptors are set-size, span, and variance of gaps between consecutive set angles.Source: §2.3 · p. 2
3. EXPERIMENTS
The experiments evaluate the tracker on simulated moving-speaker scenes with and without directional burst noise. Smoothing has its clearest reported benefit under burst noise, reducing MAE across reverberation conditions while preserving the source trajectory.
3. EXPERIMENTS
- The dynamic evaluation set contains 50 rooms with a speaker continuously moving along a 1.5 m-radius semi-circular path, while training and calibration use static scenes from 100 rooms.The static scenes are split room-wise: 80 rooms for training and 20 for calibration.Source: §3 · p. 2
- Test bursts are directional, occur at 0 dB local SNR, last 300–500 ms, and are injected 1–5 times from previously visited source positions.Simulations use five-microphone linear arrays in randomized rooms with RT60 values of 0.2, 0.5, or 0.8 seconds.Source: §3 · p. 2
- Under burst noise, smoothing consistently reduces MAE across all RT60 values; without noise, it yields minor changes and may slightly degrade low-reverberation performance.Table 1 reports before/after MAE in degrees for conditions with and without burst noise.Source: §3 · p. 2, Table 1 · p. 2
- Figure 2 illustrates a moving-speaker example in which smoothing suppresses a burst after frame 800 while retaining the circular trajectory.The conformal set expands during uncertain frames.Source: Figure 2 · p. 2
4. CONCLUSION
The paper proposes a conformal-prediction-aware DOA tracker that combines neural estimation, conformal calibration, and adaptive temporal smoothing. It reports improved localization robustness and temporal stability, particularly under burst noise.
4. CONCLUSION
- The proposed framework integrates a neural network, conformal calibration, and adaptive temporal smoothing for DOA tracking.Source: §4 · p. 2
- The framework improves localization robustness and temporal stability, particularly under burst-noise conditions.Source: §4 · p. 2
- The conclusion identifies robustness and temporal stability as the framework’s reported gains.Source: §4 · p. 2