I am using SPTK's mgcep routine to extract MCEPs from a STRAIGHT spectrogram (i.e., a spectrogram extracted from voice data by Hideki Kawahara's Maatlab STRAIGHT software, rather than a simple STFT-style spectrogram.) I encounter continual errors in the THEQ subroutine.
I understand the root cause is a failure of a check against a minimum determinant size, with a default value of 1.0e-6, and I understand that I can change that default value to make it smaller and make the error go away. What I do not know is, what effect changing this minimum value will have, or what a reasonable range of that parameter is. I have empirically determined that a value of 1.0e-38 (not a typo!) will avoid most or all of the errors, but that is 32 orders of magnitude (!!) smaller than the original default value.
Can anyone describe what the practical effects of this parameter are? Am I still likely to get results anything close to reasonable? I would be quite grateful.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The minimum value of the determinant of the normal matrix is used for only checking the sanity of the input signal. If you get the error from the mgcep command with the default value, you should check the values of the input signal (For example, if signal values are all zero in a frame, cepstral analysis cannot be performed). The minimum value of the determinant do not affect the result of cepstral analysis. It is not a parameter. I hope this answers your question.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello,
I am using SPTK's mgcep routine to extract MCEPs from a STRAIGHT spectrogram (i.e., a spectrogram extracted from voice data by Hideki Kawahara's Maatlab STRAIGHT software, rather than a simple STFT-style spectrogram.) I encounter continual errors in the THEQ subroutine.
I understand the root cause is a failure of a check against a minimum determinant size, with a default value of 1.0e-6, and I understand that I can change that default value to make it smaller and make the error go away. What I do not know is, what effect changing this minimum value will have, or what a reasonable range of that parameter is. I have empirically determined that a value of 1.0e-38 (not a typo!) will avoid most or all of the errors, but that is 32 orders of magnitude (!!) smaller than the original default value.
Can anyone describe what the practical effects of this parameter are? Am I still likely to get results anything close to reasonable? I would be quite grateful.
Hi John,
The minimum value of the determinant of the normal matrix is used for only checking the sanity of the input signal. If you get the error from the mgcep command with the default value, you should check the values of the input signal (For example, if signal values are all zero in a frame, cepstral analysis cannot be performed). The minimum value of the determinant do not affect the result of cepstral analysis. It is not a parameter. I hope this answers your question.