I've got this warning when trying to run SphinxTrain on my speech data
Current Overall Likelihood Per Frame = -0.43787555858883
Convergence ratio = -1.32930832204082 WARNING: NEGATIVE CONVERGENCE RATIO! CHECK YOUR DATA AND TRASNCRIPTS
Likelihoods have converged! Baum Welch training completed!
WHat does negative convergence ratio mean? What is the effect of it on the training accuracy?
I've got some other errors as well. Such as this one:
utt> 102 IND005_F_S_C_N_0135 1472 0 188 25 WARNING: "gauden.c", line 1370: Scaling factor too small: -733.560857
ERROR: "backward.c", line 1020: alpha(9.891610e-03) <> sum of alphas * betas (0.000000e+00) in frame 1036
ERROR: "baum_welch.c", line 331: IND005_F_S/IND005_F_S_C_N_0135 ignored
Are they related? I mean is it possible that negative convergence ratio has caused these errors?
What can I do to get rid of this warning and errors?
Thanks in advance
Eka
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Doing a search showed many other instances where this was answered so many times before.
Could be a number of reasons.
-Your transcription does not align with your wav file.
-There could be an sudden sound change in the wav.
-Senomes could be set too high for too little amount of data.
-Your segmentations are too long(10-20 seconds ideal)
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi
I've got this warning when trying to run SphinxTrain on my speech data
Current Overall Likelihood Per Frame = -0.43787555858883
Convergence ratio = -1.32930832204082
WARNING: NEGATIVE CONVERGENCE RATIO! CHECK YOUR DATA AND TRASNCRIPTS
Likelihoods have converged! Baum Welch training completed!
WHat does negative convergence ratio mean? What is the effect of it on the training accuracy?
I've got some other errors as well. Such as this one:
utt> 102 IND005_F_S_C_N_0135 1472 0 188 25 WARNING: "gauden.c", line 1370: Scaling factor too small: -733.560857
ERROR: "backward.c", line 1020: alpha(9.891610e-03) <> sum of alphas * betas (0.000000e+00) in frame 1036
ERROR: "baum_welch.c", line 331: IND005_F_S/IND005_F_S_C_N_0135 ignored
Are they related? I mean is it possible that negative convergence ratio has caused these errors?
What can I do to get rid of this warning and errors?
Thanks in advance
Eka
Hi
Doing a search showed many other instances where this was answered so many times before.
Could be a number of reasons.
-Your transcription does not align with your wav file.
-There could be an sudden sound change in the wav.
-Senomes could be set too high for too little amount of data.
-Your segmentations are too long(10-20 seconds ideal)