[hts-users:02530] About training in the STRAIGHT version HTS 2.1.1 script
- Subject: [hts-users:02530] About training in the STRAIGHT version HTS 2.1.1 script
- From: evan huang <vinsanity1101@xxxxxxxxx>
- Date: Wed, 23 Jun 2010 18:12:01 +0800
- Delivered-to: hts-users@xxxxxxxxxxxxxxx
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=domainkey-signature:mime-version:received:received:date:message-id :subject:from:to:content-type; bh=1EB5tsEqmTCKIAhjFYyTnlsYymk1lnhgzyJaJbHi3ak=; b=NRONbgAFghRBfJhQRVZ2djxS2GNqYHuL5IXuIw5+VNLagjnk+4i5zYvSkexjOwBOyb 6AUBF8qTvS7cJe1dO2L2ybiM3ohrc+/ZuEqOd8+FXwzqV0TbrPjO5AsmeDEbKK5g6RP0 nKhZJUznLJDK0LMCJ+xMV9Er38Cq+/nhe3B3o=
- Domainkey-signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=mime-version:date:message-id:subject:from:to:content-type; b=glWWRScqA4ToOMZAl+znQPgyTj+aN62nIrYF4rEwn1UqINMJIAVv8eMpGjxCjEFX9s EeUZsuo8EPyIgMeXwke/U2qq0gdvYS3Q9WYwudbesB8aJKo6GQGvDS2SxKIitzkksEtz cOoUMf7dfDV0H5XrvoeeJVvDLUbEfUSbAFUTM=
Hello,
I've downloaded the STRAIGHT version English demo script from the website and tried to train a Chinese acoustic model.
the script have been modified in order to use straightV40_007c (from 1024 to 2048, which is suggested from the mailing list) and Chinese dataset.
However, the training process stops at the "generating speech parameter sequences (stc)" and the error is showed as below:
------------------------------------------------------------------
...
WARNING [-9999] OutProb: output prob at frame 80 is 0.0 in /usr/local/HTS-2.1.1/bin/HMGenS
WARNING [-9999] OutProb: output prob at frame 81 is 0.0 in /usr/local/HTS-2.1.1/bin/HMGenS
WARNING [-9999] OutProb: output prob at frame 82 is 0.0 in /usr/local/HTS-2.1.1/bin/HMGenS
WARNING [-9999] OutProb: output prob at frame 83 is 0.0 in /usr/local/HTS-2.1.1/bin/HMGenS
.....
WARNING [-9999] ParamGen: failed to compute output prob in /usr/local/HTS-2.1.1/bin/HMGenS
.....
Pruning-On[1500.0 100.0 5000.0]
Generating Label sen_0001.lab
Total number of frames = 1377
PdfStream[1]: 1377 frames
PdfStream[2]: 793 frames
PdfStream[3]: 1377 frames
EM-based parameter generation (Hidden: MIX, STATE)
WARNING [-7324] StepBack: Bad data or over pruning
in /usr/local/HTS-2.1.1/bin/HMGenS
WARNING [-9999] ParamGen: failed to compute output prob in //usr/local/HTS-2.1.1/bin/HMGenS
Iteration 1: Average LogP = -7.262164e+06
...
-----------------------------------------------------------
Although the 1mix model can be trained, and the model can be used to generate synthesized speech.
But why the stc step has such an error?
I've used the non-straight version and it can complete the all training procedures without error using the same dataset.
Maybe it is due the STRAIGHT version?
Could anyone suggest that the suitable version of STRAIGHT that I should use? Or other useful suggestions such as the parameter setting?
Thanks very much in advance for any kinds of help!
Best Regards,
YC Huang
- Follow-Ups
-
- [hts-users:02531] Re: About training in the STRAIGHT version HTS 2.1.1 script, 康世胤