Log:Kaldi Tele 20120720
来自cslt Wiki
在Log:Kaldi_Tele_20120719中所说的情况下,结果终于能看一些了……
- monophone trainer scale: --transition-scale=1.0 --acoustic-scale=0.3 --self-loop-scale=0.3
- docoder on dev set with 4 processes
acoustic scale | WER% | time/s |
---|---|---|
0.1 | 28.94 | 1710 |
0.15 | 33.49 | 853 |
0.2 | 38.05 | 451 |
0.25 | 42.24 | 266 |
0.3 | 45.86 | 171 |
0.35 | 49.32 | 127 |
0.4 | 52.63 | 100 |
0.45 | 55.74 | 83 |
0.5 | 58.09 | 70 |
试着还原一下训练参数,加上试试更低的scale的结果
- monophone trainer scale: --transition-scale=1.0 --acoustic-scale=0.1 --self-loop-scale=0.1
- docoder on dev set with 4 processes
acoustic scale | WER% | time/s |
---|---|---|
0.05 | 28.80 | 3424 |
0.1 | 30.03 | 1733 |
0.15 | 33.84 | 846 |