Comparison Of Frequency-Fusion Mechanisms For Binaural Direction-Of-Arrival Estimation For Multiple Speakers

Daniel Fejgin, Elior Hadad, Sharon Gannot, Zbynek Koldovský, Simon Doclo

Introduction

Estimating the Direction-of-Arrival (DOA) for multiple simultaneous speakers remains a significant challenge in audio signal processing, particularly for hearing aid applications. In complex acoustic environments, such as the "cocktail party" scenario, background noise and reverberation often degrade spatial cues.

Current systems typically rely on frequency fusion, where information from various frequency bins is combined to localize sources. However, standard "broadband" fusion is susceptible to dominant speakers masking others, while "narrowband" methods—estimating direction for each frequency individually—are computationally intensive and error-prone. This research addresses these limitations by proposing an intelligent grouping of frequencies based on speaker identity prior to final DOA estimation in binaural setups.

Methodology

The core of the proposed approach is an Interaural Time Difference (ITD)-based speaker-grouped frequency fusion mechanism. The pipeline consists of four primary stages:

  1. Spatial Spectra Construction: Frequency-dependent spatial spectra (SPS) are generated using three established algorithms: Multiple Signal Classification (MUSIC), Steered Response Power (SRP), and Relative Transfer Function (RTF) vector-matching.
  2. Frequency Selection: A Coherent-to-Diffuse Ratio (CDR) criterion is applied to filter out frequency bins dominated by noise or excessive reverberation.
  3. ITD Grouping: Exploiting the consistency of ITDs for a speaker at a fixed angle, a grid-search identifies the most probable ITDs for $J$ speakers. Each time-frequency bin is then assigned to the most likely corresponding speaker.
  4. Fusion: Spatial spectra are pooled exclusively within these speaker-specific groups to derive the final DOA estimates.

This "online" grouping method is compared against a "batch" approach utilizing Independent Vector Extraction (IVE), which separates audio sources before DOA estimation.

Comparison of standard DOA estimation (left) and IVE-based source separation (right) architectures.

Experimental Setup

The methodology was evaluated using the BRUDEX database, specifically designed for binaural research. The experimental conditions included:

  • Environments: Three levels of reverberation (low, medium, and high).
  • Sources: Two static speakers (one male, one female) amidst diffuse "babble" noise.
  • Scope: 132 unique speaker-location combinations across various Signal-to-Noise Ratios (SNRs).

Results and Discussion

Experimental results indicate that the speaker-grouped approach significantly outperforms traditional fusion methods.

Localization accuracy across different frequency-fusion mechanisms (narrowband, broadband, ITD-based, and IVE-based) for MUSIC, SRP, and RTF-based methods.

Key findings include:

  • Accuracy Improvements: The ITD-based grouping mechanism improved localization accuracy by at least 8% over standard broadband fusion across all tested methods.
  • Algorithm Performance: The MUSIC algorithm benefited most, showing a 12% improvement over narrowband methods.
  • Comparison with Source Separation: The proposed ITD-based method achieved performance comparable to the more complex IVE-based separation for SRP and RTF methods. Notably, it outperformed IVE by 6% when paired with the MUSIC algorithm.
  • Robustness: The use of CDR thresholds for frequency selection consistently enhanced results, particularly for subspace-based methods like MUSIC.

By grouping frequencies by ITD first, the system effectively mitigates the "dominant speaker" bias, preventing loud sources from masking quieter ones or creating artifacts in the spatial map.

Conclusion

This study demonstrates that ITD-based frequency grouping provides a robust and efficient intermediate step for multi-speaker localization in binaural systems. While the current model assumes a known number of speakers and static positions, it establishes a strong foundation for real-time hearing aid applications. Future research will focus on automatic speaker count detection and adapting the grouping mechanism for dynamic, moving sources in three-dimensional space.

Summary updated

Abstract

To estimate the direction of arrival (DOA) of multiple speakers with methods that use prototype transfer functions, frequency-dependent spatial spectra (SPS) are usually constructed. To make the DOA estimation robust, SPS from different frequencies can be combined. According to how the SPS are combined, frequency fusion mechanisms are categorized into narrowband, broadband, or speaker-grouped, where the latter mechanism requires a speaker-wise grouping of frequencies. For a binaural hearing aid setup, in this paper we propose an interaural time difference (ITD)-based speaker-grouped frequency fusion mechanism. By exploiting the DOA dependence of ITDs, frequencies can be grouped according to a common ITD and be used for DOA estimation of the respective speaker. We apply the proposed ITD-based speaker-grouped frequency fusion mechanism for different DOA estimation methods, namely the multiple signal classification, steered response power and a recently published method based on relative transfer function (RTF) vectors. In our experiments, we compare DOA estimation with different fusion mechanisms. For all considered DOA estimation methods, the proposed ITD-based speaker-grouped frequency fusion mechanism results in a higher DOA estimation accuracy compared with the narrowband and broadband fusion mechanisms.
BibTeX
@inproceedings{fejgin2024comparison,
  author = {Daniel Fejgin and Elior Hadad and Sharon Gannot and Zbynek Koldovský and Simon Doclo},
  title = {{Comparison Of Frequency-Fusion Mechanisms For Binaural Direction-Of-Arrival Estimation For Multiple Speakers}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
  url = {https://openreview.net/forum?id=PqdEAvvDRL},
}
Figure