Multichannel source counting with CRNN : analysis of the performance

Pierre-Amaury Grumiaux (1) , Srdan Kitic (2, 1) , Laurent Girin (3) , Alexandre Guérin (1)

1 Orange Labs
2 PANAMA - Parcimonie et Nouveaux Algorithmes pour le Signal et la Modélisation Audio
3 GIPSA-CRISSP - GIPSA - Cognitive Robotics, Interactive Systems, & Speech Processing

In this work we focus on the problem of estimating the number of concurrent speaker in an audio recording. This information is often a prerequisite in several audio processing tasks such as speaker separation, localization and tracking. In a previous work, we proposed to tackle this problem by using a convolutional recuurrent neural network (CRNN) with first-order Ambisonics input features. The network was trained to predict up to 5 concurrent speaker with a simulated dataset which includes many different conditions in terms of source and microphone positions, reverberation and noise. In this work, we analyze the performance of the neural network along the frames of an input signal. We show that there is an optimal analysis frame within the sequence for which the performance is better, and that it depends on some hyperparameters of the network, such as number of convolutional layers, convolutional kernel sizes, or number of timesteps in the recurrent part. This provides a good insight into the behavior of CRNN on audio signals for this specific task.

Keywords: source counting; convolutional recurrent neural network; first-order Ambisonics; spatial room impulse response

Domains: Acoustics [physics.class-ph]; Vibrations [physics.class-ph]

Pierre-Amaury Grumiaux, Srdan Kitic, Laurent Girin, Alexandre Guérin. Multichannel source counting with CRNN : analysis of the performance. Forum Acusticum 2020, Dec 2020, Lyon (virtual), France. pp.829-835, ⟨10.48465/fa.2020.0766⟩. ⟨hal-03235360⟩