Multi-Source Direction of Arrival Estimation of Noisy Speech using Convolutional Recurrent Neural Networks with Higher-Order Ambisonics Signals

Convolutional recurrent neural networks provide state of the art results in direction of arrival estimation based on first-order Ambisonics signals, especially in the presence of noise and/or interfering sound sources. In this work, we investigate whether increasing the order of Ambisonics up to the fourth order further improves the estimation results in a challenging multi-speaker setting with two or three simultaneously active speakers. Our results show that each additional order of the Ambisonics representation further improves the localization performance for both speech signals based on simulated and real measured spatial room impulse responses. The greatest gains in accuracy can be observed in the particularly demanding scenarios with three speakers and poor signal-to-interference-ratio.

I. Introduction

Direction of arrival (DOA) estimation is a well-known task in audio signal processing and an important component of many applications such as speech separation [1] or speech enhancement [2]. Neural networks have been shown to be superior to classical parametric approaches in DOA estimation, especially in very demanding reverberant, noisy, and low-SNR environments [3]–[6]. In addition, Ambisonics-based audio signal processing is becoming increasingly popular due to the flexibility and generalizability it enables. Therefore, DOA estimation based on first-order Ambisonics (FOA) signals has been the subject of much attention [6]–[9].