Localizing audio sources is challenging in real reverberant environments, especially when several sources are active. We propose to use a neural network built from stacked convolutional and recurrent layers in order to estimate the directions of arrival of multiple sources from a first-order Ambisonics recording. It returns the directions of arrival over a discrete grid of a known number of sources. We propose to use features derived from the acoustic intensity vector as inputs. We analyze the behavior of the neural network by means of a visualization technique called layerwise relevance propagation. This analysis highlights which parts of the input signal are relevant in a given situation. We also conduct experiments to evaluate the performance of our system in various environments, from simulated rooms to real recordings, with one or two speech sources. The results show that the proposed features significantly improve performances with respect to raw Ambisonics inputs.

We present a source localization system for first-order Ambisonics (FOA) contents based on a stacked convolutional and recurrent neural network (CRNN). We propose to use as input to the CRNN the FOA acoustic intensity vector, which is easy to compute and closely linked to the sound direction of arrival (DoA). The system estimates the DoA of a point source in both azimuth and elevation. We conduct an experimental evaluation in configurations including reverberation, noise, and various speaker w.r.t. microphone orientations. The results show that the proposed architecture and input allow the network to return accurate location estimates in realistic conditions compared to another recent CRNN-based system.

We compare the performance of regression and classification neural networks for single-source direction-of-arrival estimation. Since the output space is continuous and structured, regression seems more appropriate. However, classification on a discrete spherical grid is widely believed to perform better and is predominantly used in the literature. For regression, we propose two ways to account for the spherical geometry of the output space based either on the angular distance between spherical coordinates or on the mean squared error between Cartesian coordinates. For classification, we propose two alternatives to the classical one-hot encoding framework: we derive a Gibbs distribution from the squared angular distance between grid points and use the corresponding probabilities either as soft targets or as cross-entropy weights that retain a clear probabilistic interpretation. We show that regression on Cartesian coordinates is generally more accurate, except when localized interference is present, in which case classification appears to be more robust.

No abstract

We present a CNN architecture for speech enhancement from multichannel first-order Ambisonics mixtures. The data-dependent spatial filters, deduced from a mask-based approach, are used to help an automatic speech recognition engine to face adverse conditions of reverberation and competitive speakers. The mask predictions are provided by a neural network, fed with rough estimations of speech and noise amplitude spectra, under the assumption of known directions of arrival. This study evaluates the replacing of the recurrent LSTM network previously investigated by a convolutive U-net under more stressing conditions with an additional second competitive speaker. We show that, due to more accurate short-term masks prediction, the U-net architecture brings some improvements in terms of word error rate. Moreover, results indicate that the use of dilated convolutive layers is beneficial in difficult situations with two interfering speakers, and/or where the target and interferences are close to each other in terms of the angular distance. Moreover, these results come with a two-fold reduction in the number of parameters.