Fig. 8: Real-world evaluations with cluttered surfaces and human speakers.
From: Creating speech zones with self-distributing acoustic swarms
We assess our system’s performance in three previously unseen cluttered environments, shown in (A–C). In (D), we show precision and recall results by varying the number of speakers in the audio mixture. E illustrates the cumulative distribution function (CDF) for the 2D localization errors across all three environments, considering varying numbers of speakers. F compares our approach, with and without the inter-speaker attention mechanism, and an oracle approach (IRM) (the error bar shows the standard deviation).G Shows the 2D localization errors for a participant in different locations. H Plots the mean 2D localization error as a function of the human head orientation, where 0° is when the human face is pointing in the direction of the acoustic swarm and 1800 is when the human’s back is facing the acoustic swarm. I Shows the mean 2D errors as a function of different distances of a human speaker to a wall.
