Significance of CLPSO-Based Dataset in Self-Supervised Lightweight ANN for Estimating Highly Intelligible Microphone Sensor Location

Abstract

This letter proposes training a lightweight artificial neural network (ANN) in a self-supervised manner using an optimal dataset generated via comprehensive learning particle swarm optimization (CLPSO). Although CLPSO can suggest the “optimal” microphone sensor locations in a room relative to a speaker, it is computationally taxing. Instead, we propose using these suggested sensor locations as implicit labels for training a network. It is suggested to use five-best sensor locations for training instead of one to ensure that the model captures the relationship between the speaker and the sensor locations within the room. This training is done on a resource-constrained Raspberry Pi. The trained ANN quickly predicts good sensor locations corresponding to high intelligibility in terms of short-time objective intelligibility (STOI). This performance is generalized across different combinations of room dimensions and speaker locations and is robust for varying datasets. The predictions were also validated in real-world conditions through mean opinion score (MOS) values.