Multitask Learning with Local Attention for Tibetan Speech Recognition

Table 7

Speaker ID recognition accuracy (%) of three-task models.

Architecture	Model	Lhasa-Ü-Tsang	Changdu-Kham	Amdo pastoral
SpeakerID model		67.75	93.13	95.31
WaveNet-CTC with dialect ID and speaker ID	S-D-S	72.91	98.8	96.12
	D-S-S1	70.21	95.23	93.6
	D-S-S2	70.35	96.42	96.89
Attention (5)-WaveNet-CTC	S-D-S	61.08	83.33	89.53
	D-S-S1	62.12	83.33	87.01
	D-S-S2	61.99	84.52	90.11
WaveNet-Attention (5)-CTC	S-D-S	61.99	85.71	92.05
	D-S-S1	62.53	82.14	91.08
	D-S-S2	61.18	89.28	92.44
WaveNet-Attention (7)-CTC	S-D-S	60.91	85.71	91.66
	D-S-S1	62.04	84.31	92.01
	D-S-S2	58.49	86.90	90.69
WaveNet-Attention (10)-CTC	S-D-S	58.49	84.52	92.05
	D-S-S1	59.43	83.33	91.27
	D-S-S2	63.47	92.85	97.86