Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neural Directional Filtering: Far-Field Directivity Control With a Small Microphone Array

Julian Wechsler, Srikanth Raj Chetupalli, Mhd Modar Halimeh, Oliver Thiergart, Emanuël A. P. Habets

TL;DR

This work addresses controlling far-field directivity with a compact microphone array by learning a target pattern through neural directional filtering. A lightweight FT-JNF architecture predicts a single-channel complex mask from multi-channel inputs, applying it to a reference mic to realize the desired pattern $Z_{\textrm{VDM}}$ via $\widehat{Z}_{VDM}[f,t] = \mathcal{M}[f,t]\,Y_{1}[f,t]$. The study investigates how training data composition affects pattern realization and demonstrates that the method can closely approximate cardioid and higher-order DMA patterns using few microphones, outperforming traditional parametric baselines in most cases. Results indicate strong performance when trained on multi-speaker mixtures, with best mean SDRs of $26.2$ dB for cardioid and $18.4$ dB for the $3^{\textrm{rd}}$-order DMA, suggesting practical impact for flexible spatial audio capture and reproduction. Future work includes steerable/arbitrary patterns, near-field and reverberant scenarios, measured-data validation, and exploring VDM placement strategies.

Abstract

Capturing audio signals with specific directivity patterns is essential in speech communication. This study presents a deep neural network (DNN)-based approach to directional filtering, alleviating the need for explicit signal models. More specifically, our proposed method uses a DNN to estimate a single-channel complex mask from the signals of a microphone array. This mask is then applied to a reference microphone to render a signal that exhibits a desired directivity pattern. We investigate the training dataset composition and its effect on the directivity realized by the DNN during inference. Using a relatively small DNN, the proposed method is found to approximate the desired directivity pattern closely. Additionally, it allows for the realization of higher-order directivity patterns using a small number of microphones, which is a difficult task for linear and parametric directional filtering.

Neural Directional Filtering: Far-Field Directivity Control With a Small Microphone Array

TL;DR

This work addresses controlling far-field directivity with a compact microphone array by learning a target pattern through neural directional filtering. A lightweight FT-JNF architecture predicts a single-channel complex mask from multi-channel inputs, applying it to a reference mic to realize the desired pattern via . The study investigates how training data composition affects pattern realization and demonstrates that the method can closely approximate cardioid and higher-order DMA patterns using few microphones, outperforming traditional parametric baselines in most cases. Results indicate strong performance when trained on multi-speaker mixtures, with best mean SDRs of dB for cardioid and dB for the -order DMA, suggesting practical impact for flexible spatial audio capture and reproduction. Future work includes steerable/arbitrary patterns, near-field and reverberant scenarios, measured-data validation, and exploring VDM placement strategies.

Abstract

Capturing audio signals with specific directivity patterns is essential in speech communication. This study presents a deep neural network (DNN)-based approach to directional filtering, alleviating the need for explicit signal models. More specifically, our proposed method uses a DNN to estimate a single-channel complex mask from the signals of a microphone array. This mask is then applied to a reference microphone to render a signal that exhibits a desired directivity pattern. We investigate the training dataset composition and its effect on the directivity realized by the DNN during inference. Using a relatively small DNN, the proposed method is found to approximate the desired directivity pattern closely. Additionally, it allows for the realization of higher-order directivity patterns using a small number of microphones, which is a difficult task for linear and parametric directional filtering.
Paper Structure (14 sections, 6 equations, 3 figures, 2 tables)

This paper contains 14 sections, 6 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Proposed Method
  4. Architecture
  5. Loss function
  6. Training strategy and simulations
  7. Experimental Setup
  8. Virtual microphone directivity patterns
  9. Array Geometry and Datasets
  10. Training details
  11. Performance Evaluation
  12. Performance metrics
  13. Performance analysis
  14. Conclusion

Figures (3)

  • Figure 1: Logarithmic polar plots of the DMA patterns considered in this study.
  • Figure 2: Distributions of SDR values when testing the best model for (a) the cardioid target and (b) the $3^{\textrm{rd}}$-order DMA target. The models were tested with two concurrently active sources. Their DOA are given on the axes, SDR for missing DOA combinations were interpolated using cubic interpolation.
  • Figure 3: Polar plot of the two VDM targets and their corresponding DNN estimate. The gray area illustrates the standard deviation.