Realization of a Model-Agnostic Explainable AI Method for Audio Classification

Overview

This repository introduces a novel, model-agnostic Explainable AI (XAI) framework tailored specifically for audio classification tasks. It features three specialized architectures: RISE-SPEC, RISE-WAVE, and RISE-AUDIO which consistently outperform traditional baseline XAI methods, including RISE, LIME, and Grad-CAM.

XAI Evaluation on ESC-50

Spectrogram-based Models (2D)

Method		ResNet50			HTS-AT
	Ins(%)↑	Del(%)↓	OA(%)↑	Ins(%)↑	Del(%)↓	OA(%)↑
LIME	78.90	29.50	49.40	63.28	26.45	36.83
Grad-CAM	76.25	30.69	45.56	--	--	--
RISE (Original)	80.63	31.13	49.50	56.61	24.05	32.56
RISE-SPEC	82.02	22.26	59.76	57.26	19.91	37.35
RISE-AUDIO	75.88	20.58	55.30	74.29	36.72	37.57

Waveform-based Models (1D)

Method		ACDNet			Wav2Vec2
	Ins(%)↑	Del(%)↓	OA(%)↑	Ins(%)↑	Del(%)↓	OA(%)↑
LIME	47.31	15.67	31.64	71.34	24.62	46.72
RISE (Original)	51.80	27.91	23.89	70.35	33.53	36.82
RISE-WAVE	60.03	32.28	27.75	78.54	40.45	38.09
RISE-AUDIO	77.39	6.99	70.40	92.53	24.96	67.57