Multitask Learning with Local Attention for Tibetan Speech Recognition

Table 1

The experimental data statistics.

Dialect	Training data (hours)	Training utterances	Test data (hours)	Test utterances	Speaker
Lhasa-Ü-Tsang	4.40	6678	0.49	742	20
Changdu-Kham	1.90	3004	0.19	336	6
Amdo pastoral	3.28	4649	0.37	516	14
Total	9.58	14331	1.05	2110	40