Tanel Alumäe - 2006-11-13

Hello,

I noticed that there are many errors as follows in the output of bw when executing the training:

utt> 19273 m_50498_2661_0045 626 0 208 37 WARNING: "gauden.c", line 1380: Scaling factor too small: -1
525.342324
ERROR: "backward.c", line 1108: alpha(3.330182e-03) <> beta(0.000000e+00)
ERROR: "backward.c", line 1115: beta underflow
5 8 3.322895e-42ERROR: "baum_welch.c", line 304: 2661/m_50498_2661_0045 ignored

and:

utt> 19276 m_50498_2661_0048 1120 0 112 23 WARNING: "gauden.c", line 1380: Scaling factor too small: -1
308.806708
ERROR: "backward.c", line 1108: alpha(8.619176e-02) <> beta(6.959324e-267)
5 9 1.733918e-42ERROR: "baum_welch.c", line 304: 2661/m_50498_2661_0048 ignored

In all such cases, the reported beta is very small (or 0)

Interestingly, when I turn on variance normalization, the errors are a bit different:

utt> 16325 m_50387_3900_0030 803 0 44 43 WARNING: "gauden.c", line 1380: Scaling factor too small: -7
46.691241
ERROR: "backward.c", line 965: alpha(5.000000e-01) <> sum of alphas * betas (5.366818e-22) in frame 536
ERROR: "baum_welch.c", line 304: 3900/m_50387_3900_0030 ignored

I listened to some of such utterances and they do match with the transcripts.

The speech that I'm training with contains short utterances from a wide variety of speakers, recorded over telephone.
I don't get such errors when training on another database that contains very clean speech, recorded in a studio.

I'm using standard 1s_c_d_dd features, and latest sphinx from SVN.