Speaker adaptation based on regularized speaker-dependent eigenphone matrix estimation

EURASIP Journal on Audio, Speech, and Music Processing

Table 5 Average tonal syllable recognition rate (%) after eigenphone-based speaker adaptation using sparse group lasso

λ ₃	Number of adaptation sentences
	1	2	4	6	8	10
10	53.78	56.57	58.14	59.06	60.05	60.91
	(0.61, 0.01)	(0.47, 0.0)	(0.31, 0.0)	(0.22, 0.0)	(0.18, 0.0)	(0.15, 0.0)
20	54.76	56.74	58.29	59.21	60.18	60.93
	(0.62, 0.01)	(0.45, 0.0)	(0.31, 0.0)	(0.22, 0.0)	(0.18, 0.0)	(0.15, 0.0)
30	54.55	56.86	58.55	59.53	60.20	61.25
	(0.63, 0.02)	(0.44, 0.0)	(0.32, 0.0)	(0.23, 0.0)	(0.18, 0.0)	(0.15, 0.0)
40	54.49	56.65	58.35	59.32	60.11	60.93
	(0.63, 0.05)	(0.43, 0.0)	(0.31, 0.0)	(0.23, 0.0)	(0.18, 0.0)	(0.16, 0.0)
80	54.13	56.04	57.72	58.92	59.90	60.43
	(0.78, 0.37)	(0.45, 0.02)	(0.33, 0.0)	(0.23, 0.0)	(0.19, 0.0)	(0.16, 0.0)
120	54.05	54.95	57.01	58.35	59.38	60.24
	(0.91, 0.76)	(0.58, 0.21)	(0.35, 0.01)	(0.23, 0.0)	(0.19, 0.0)	(0.16, 0.0)

The number of eigenphones (N) was fixed to 100. λ₁=10.0, λ₂=0, and λ₃ was varied between 10 and 150. the average overall sparsity and column sparsity of the eigenphone matrix are shown in parentheses as pairs.