Sphinxtrain error: "All 2 active states, 90 91, zero at time 1"

[Amit Rathi]

I am training a model for female pop songs that I ran through vocal extraction using REPET. They are further cut into smaller 8 to 10 second segments for training. But the training is failing on "MODULE: 20 Training Context Independent models" stage in first forward-backward iteration itself. I get thousands of following errors in logs:

INFO: cmn.c(183): CMN: 64.51 11.57 -10.37 12.40 -1.83 -5.27 -5.79 -12.02 -21.80 1.13 -0.60 -15.25 -2.62
ERROR: "forward.c", line 594: All 2 active states, 90 91, zero at time 1
ERROR: "baum_welch.c", line 324: Adele/sample43 ignored
utt> 15803 sample43 1049 0 244 0

Things I verified:
1. Audio is 16bit 16Khz mono.
2. Audio sounds clean when heard manually.
3. Text and audio matches each other.
4. Training has worked fine when I used KAML vocal extraction and even training on songs directly. So there is no problem in my installation of sphinxbase/sphinxtrain.

Any inputs?

Ref:
REPET: http://zafarrafii.com/repet.html#REPET-SIM

[Nickolay V. Shmyrev]

You might share the audio after repet.

[Amit Rathi]

That clip is after REPET extraction. Here are couple of more samples for your analysis.

[Amit Rathi]

Any updates?