Multitask Learning with Local Attention for Tibetan Speech Recognition

Table 6

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

Architecture	Model	Lhasa-Ü-Tsang	Changdu-Kham	Amdo Pastoral
DialectID model		97.88	92.24	97.9
WaveNet-CTC with dialect ID and speaker ID	D-S-S1	98.01	98.8	99.41
	D-S-S2	99.73	96.42	99.61
	S-D-S	99.25	95.23	99.03
Attention (5)-WaveNet-CTC	S-D-S	100	76.19	91.27
	D-S-S1	100	90.47	94.18
	D-S-S2	100	82.14	93.02
WaveNet-Attention (5)-CTC	S-D-S	100	89.28	93.79
	D-S-S1	100	85.71	93.79
	D-S-S2	100	95.23	94.18
WaveNet-Attention (7)-CTC	S-D-S	0	85.71	91.66
	D-S-S1	0	89.98	93.88
	D-S-S2	0	89.28	95.34
WaveNet-Attention (10)-CTC	S-D-S	0	85.71	95.54
	D-S-S1	0	94.04	93.99
	D-S-S2	0	0	0