An Experimental Study on SALSA-Lite Method for Localization of Multiple Human Speakers

SALSA-Lite is a fast and efficient feature designed for localization and detection of polyphonic sound events (SELD) with microphone array inputs. It serves as a lightweight version of the original SALSA feature for polyphonic SELD. In this paper, we specifically use a SELD network based on SALSA-Lite features to localize sound events from multiple human female/male speakers. We have conducted a series of experiments to evaluate its effectiveness in determining Direction of Arrival (DoA) solely based on auditory events derived from speech. While the base method's superior performance regarding complexity and accuracy has been established in the localization of sound events relative to leading methodologies, we show that it encounters additional difficulties when speech sources are in close proximity to one another, particularly when noise and reverberation significantly obscure speech signals. We also demonstrate that the number of active speakers, their distances from the microphone array, and their tone of voice can significantly impact the performance of SELD network based on SALSA-Lite features.