Learning Deep Embedding with Acoustic and Phoneme Features for Speaker Recognition in FM Broadcasting

Table 2

TI-SV results for the hybrid DNNs compared with existing benchmarks.

Feature	Method	Aggregation	Loss	EER (%)
AC	Nagrani et al. [11]	—	—	8.8
	Nagrani et al. [11]	TAP	Softmax	10.2
	Kim and Park [26]	TAP	AAM-Softmax	5.68
	Han et al. [27]	SAP	Softmax	5.75
	Han et al. [27]	SAP	AM-Softmax	4.15
	Cai et al. [28]	SAP	A-Softmax	4.40
	Cai et al. [28]	LDE	A-Softmax	4.48
	Wang et al. [29]	MHA	CosAMS	4.46
	Wang et al. [29]	MRMHA	CosAMS	4.10
	Wang et al. [29]	MRMHA	CosAMS	3.98
	Wang et al. [29]	MRMHA	CosAMS	3.96
AC&PH	ours_PFI_1	TAP	AAM-Softmax	4.24
	ours_PFI_2	TAP	AAM-Softmax	4.46
	ours_PFI_1	AWP	AAM-Softmax	3.72
	ours_PFI_2	AWP	AAM-Softmax	3.84

AC, acoustic; PH, phoneme; TAP, temporal average pooling; SAP, self-attention pooling; LDE, learnable dictionary encoding; MHA, multihead attention; MRMHA, multiresolution multihead attention; AWP, adaptive weight pooling. The use of “bold” is to emphasize the experimental result (3.72%).