Tracking Multiple Audio Sources with the von Mises Distribution and Variational EM

Yutong Ban, Xavier Alameda-Pineda, Christine Evers, Radu Horaud

Abstract

In this paper, we address the problem of simultaneously tracking several moving audio sources, namely the problem of estimating source trajectories from a sequence of observed features. We propose to use the von Mises distribution to model audio-source directions of arrival with circular random variables. This leads to a Bayesian filtering formulation which is intractable because of the combinatorial explosion of associating observed variables with latent variables, over time. We propose a variational approximation of the filtering distribution. We infer a variational expectation-maximization algorithm that is both computationally tractable and time efficient. We propose an audio-source birth method that favors smooth source trajectories and which is used both to initialize the number of active sources and to detect new sources. We perform experiments with the recently released LOCATA dataset comprising two moving sources and a moving microphone array mounted onto a robot.

1. Introduction

Tracking multiple moving audio sources is a fundamental challenge for technologies such as home assistants, robot audition, and speech enhancement. Unlike tracking objects in a 2D Euclidean plane, audio tracking often relies on Directions of Arrival (DOAs)—angles representing the sound source relative to a microphone array. This modality presents three major hurdles: the circularity of angular data, the data association problem (linking specific observations to specific sources), and the resulting combinatorial explosion that makes traditional Bayesian filtering computationally prohibitive for real-time applications.

2. Proposed Methodology

2.1 The von Mises Distribution

Standard tracking models often utilize Gaussian distributions, which fail to account for the periodic nature of azimuth angles (where 0° and 360° are identical). We propose using the von Mises distribution, the circular equivalent of a Normal distribution. By modeling DOAs as circular random variables, our framework naturally handles boundary conditions without the errors inherent in linear approximations.

2.2 Variational Expectation-Maximization (VEM)

To overcome the intractability of the association problem, we employ a variational approximation. By assuming that source locations and observation-to-source assignments are conditionally independent, we derive a computationally efficient Variational Expectation-Maximization (VEM) algorithm:

  • E-Step: The system estimates the probability of each observation belonging to a specific source and updates the estimated location of each active speaker.
  • M-Step: The model parameters, including noise levels and source dynamics, are updated to maximize the likelihood of the observed data.

2.3 Audio-Source Birth Process

To handle dynamic environments where speakers enter or leave, we implement a birth process based on the "harmonic sum theorem." The system monitors unassigned DOAs; if a sequence of observations shows high statistical consistency over a short temporal window, a new track is initialized. This prevents "ghost" tracks caused by momentary noise while ensuring rapid detection of new sources.

3. Experimental Evaluation

3.1 Dataset and Setup

We evaluated our method using Task 6 of the LOCATA (LOCalization and Tracking of Acoustic sources) dataset. This task is particularly challenging as it involves two moving speakers and a moving microphone array mounted on a biped humanoid robot, introducing significant ego-noise, reverberation, and real-world acoustic complexity.

3.2 Quantitative Results

The proposed vM-VEM algorithm was compared against several state-of-the-art trackers, including the von Mises Probability Hypothesis Density (vM-PHD) filter and Gaussian-based models (GM-ZO and GM-FO).

Performance comparison of the proposed vM-VEM against state-of-the-art methods.

As shown in the table above, vM-VEM achieved a Mean Absolute Error (MAE) of 2.6°, the lowest among all tested methods. Furthermore, it maintained a superior False Alarm (FA) rate of 5.9%, demonstrating its robustness against noise compared to the 9.5%–10.8% rates seen in competing models.

3.3 Qualitative Analysis

Qualitatively, the vM-VEM produces significantly smoother and more consistent trajectories than the PHD filter. While PHD filters often struggle with track continuity and explicit source labeling without heavy post-processing, the VEM framework provides explicit labels and maintains track stability even during complex movements.

Estimated trajectories for vM-PHD, GM-FO, and the proposed vM-VEM compared to ground-truth.

4. Discussion and Conclusion

Our results demonstrate that circular statistics are vital for audio tracking; even with a simple zero-order dynamic model, the von Mises approach outperformed higher-order Gaussian models. The variational factorization proved to be the "aha moment" for achieving tractability, allowing for real-time performance without sacrificing accuracy.

Future work will focus on incorporating first-order dynamics (velocity) to better handle fast-moving sources and improving track retention during long periods of source silence. Scaling the VEM approach to handle dozens of simultaneous sources in highly reverberant environments remains a promising direction for future research.

Summary updated

Abstract

In this paper we address the problem of simultaneously tracking several moving audio sources, namely the problem of estimating source trajectories from a sequence of observed features. We propose to use the von Mises distribution to model audio-source directions of arrival with circular random variables. This leads to a Bayesian filtering formulation which is intractable because of the combinatorial explosion of associating observed variables with latent variables, over time. We propose a variational approximation of the filtering distribution. We infer a variational expectation-maximization algorithm that is both computationally tractable and time efficient. We propose an audio-source birth method that favors smooth source trajectories and which is used both to initialize the number of active sources and to detect new sources. We perform experiments with the recently released LOCATA dataset comprising two moving sources and a moving microphone array mounted onto a robot.
BibTeX
@article{ban2019tracking,
  author = {Yutong Ban and Xavier Alameda-Pineda and Christine Evers and Radu Horaud},
  title = {{Tracking Multiple Audio Sources with the von Mises Distribution and Variational EM}},
  journal = {IEEE Signal Processing Letters},
  year = {2019},
  url = {https://openreview.net/forum?id=CpbCX9A_qZ3},
}
Figure