Multiclass audio segmentation based on recurrent neural networks for broadcast domain data

EURASIP Journal on Audio, Speech, and Music Processing

Table 2 SER, error per class and average class error for the 1BLSTM RNN classifier on the test partition for different frontend configurations (Chr, chroma; Δ,ΔΔ, 1st and 2nd order derivatives)

Feats	SER	Class error (%)				Avg
		mu	sp	sm	sn
64 bands	18.18	18.54	32.43	32.48	35.76	29.80
80 bands	17.70	18.19	31.33	31.41	34.91	28.96
96 bands	17.93	20.68	30.84	32.09	34.25	29.46
64 bands + Chr	16.97	18.83	30.88	29.92	32.76	28.10
80 bands + Chr	17.89	19.77	32.23	29.55	33.92	28.87
96 bands + Chr	17.65	19.75	30.68	31.62	33.66	28.93
64 bands + Chr + Δ,ΔΔ	16.61	17.46	29.93	29.26	32.60	27.31
80 bands + Chr + Δ,ΔΔ	16.25	16.82	30.00	26.75	32.07	26.41
96 bands + Chr + Δ,ΔΔ	16.46	17.38	29.92	27.98	32.70	27.00