Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DNN-Based Online Source Counting Based on Spatial Generalized Magnitude Squared Coherence

Henri Gode, Simon Doclo

TL;DR

The paper addresses online counting of active sound sources in reverberant, multichannel settings by re-framing source counting as a change-detection problem using spatial coherence. It introduces activation features based on whitening and generalized magnitude-squared coherence and deactivation features via time-reversed whitening, then replaces threshold-based counts with DNN-based classifiers that take concatenated activation/deactivation features to estimate the per-frame source count $K_t$. Two causal architectures (TCN and RNN) are trained to output a softmax over possible counts and are evaluated on BRUDEX/Librispeech-derived binaural data with up to four sources; results show significant gains, notably when including deactivation features (e.g., 91.9% frame accuracy). The approach is real-time capable (factor ~0.007) and improves robustness over the conventional activation-only thresholding, enabling more reliable online source counting for downstream tasks such as localization, separation, and speech enhancement in hearing aids.

Abstract

The number of active sound sources is a key parameter in many acoustic signal processing tasks, such as source localization, source separation, and multi-microphone speech enhancement. This paper proposes a novel method for online source counting by detecting changes in the number of active sources based on spatial coherence. The proposed method exploits the fact that a single coherent source in spatially white background noise yields high spatial coherence, whereas only noise results in low spatial coherence. By applying a spatial whitening operation, the source counting problem is reformulated as a change detection task, aiming to identify the time frames when the number of active sources changes. The method leverages the generalized magnitude-squared coherence as a measure to quantify spatial coherence, providing features for a compact neural network trained to detect source count changes framewise. Simulation results with binaural hearing aids in reverberant acoustic scenes with up to 4 speakers and background noise demonstrate the effectiveness of the proposed method for online source counting.

DNN-Based Online Source Counting Based on Spatial Generalized Magnitude Squared Coherence

TL;DR

The paper addresses online counting of active sound sources in reverberant, multichannel settings by re-framing source counting as a change-detection problem using spatial coherence. It introduces activation features based on whitening and generalized magnitude-squared coherence and deactivation features via time-reversed whitening, then replaces threshold-based counts with DNN-based classifiers that take concatenated activation/deactivation features to estimate the per-frame source count . Two causal architectures (TCN and RNN) are trained to output a softmax over possible counts and are evaluated on BRUDEX/Librispeech-derived binaural data with up to four sources; results show significant gains, notably when including deactivation features (e.g., 91.9% frame accuracy). The approach is real-time capable (factor ~0.007) and improves robustness over the conventional activation-only thresholding, enabling more reliable online source counting for downstream tasks such as localization, separation, and speech enhancement in hearing aids.

Abstract

The number of active sound sources is a key parameter in many acoustic signal processing tasks, such as source localization, source separation, and multi-microphone speech enhancement. This paper proposes a novel method for online source counting by detecting changes in the number of active sources based on spatial coherence. The proposed method exploits the fact that a single coherent source in spatially white background noise yields high spatial coherence, whereas only noise results in low spatial coherence. By applying a spatial whitening operation, the source counting problem is reformulated as a change detection task, aiming to identify the time frames when the number of active sources changes. The method leverages the generalized magnitude-squared coherence as a measure to quantify spatial coherence, providing features for a compact neural network trained to detect source count changes framewise. Simulation results with binaural hearing aids in reverberant acoustic scenes with up to 4 speakers and background noise demonstrate the effectiveness of the proposed method for online source counting.
Paper Structure (12 sections, 16 equations, 3 figures, 1 table)

This paper contains 12 sections, 16 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Signal Model
  3. Conventional Source Activation Detection
  4. Whitened GMSC Features for Source Activation Detection
  5. Threshold-Based Detection
  6. Source Deactivation Detection Features
  7. Proposed DNN-Based Source Count Estimation
  8. Evaluation
  9. Dataset
  10. Algorithmic Implementation and Training
  11. Results
  12. Conclusion

Figures (3)

  • Figure 1: Exemplary source activity timeline with 4 sources, showing activity intervals and the total number of active sources.
  • Figure 2: Illustration of the GMSC-based features ($=9.5dB$). Top: GMSC features for activation. Middle: GMSC features for deactivation. Bottom: Corresponding broadband features after recursive smoothing.
  • Figure 3: Acoustic setup of the BRUDEX database fejgin2023brudex.