[hts-users:04037] Re: STRAIGHT voice problem

[Martin Sulír <martin.sulir@xxxxxxxxx>]

Hi,

thanks for your reply. I checked also a binary files, and they seem to be fine. So, is It possible that there is any problem with training step? Can you think of anything else what can cause these problems?

Thanks.

MS

2014-03-08 10:08 GMT+01:00 Bajibabu Bollepalli <bajibabu7@xxxxxxxxx>:

Hi,

You can use either "x2x" command in SPTK or from NumPy in python to read the binary files. Commands are:

" x2x +fa +a[dim] binary.file > ascii.file "
" a = numpy.fromfile("binary.file", dtype='<f4', count=-1) "

Cheers,

Bajibabu

On 8 March 2014 14:42, Martin Sulír <martin.sulir@xxxxxxxxx> wrote:

Hello,

I use this version of STRAIGHT: http://www.wakayama-u.ac.jp/~kawahara/STRAIGHTtrial/

I checked label files and they look fine, I use the same files with original HMM scirpts and they work without problem. I also checked F0 files where the number of frames and also their F0 values are ok. I checked ap, sp and there number of lines is equal to number of frames in the recording. Is there any way how to check binary cmp, mgc and bap files?

I try to train voice with my old small databse with the same paths to all tools and programs (hts, straight, sptk...) and I obtained voice without problem.

Thanks for your reply.

Regards, Martin Sulír.

2014-03-06 12:36 GMT+01:00 Sébastien Le Maguer <Sebastien.Le_maguer@xxxxxxxx>:

Hello,

which version of straight do you use ?


To find the problem,I suggest to look at the label lines containing the indicated identifier (l^k-o+u=a@2_2/A:1_1_3/B:0-0-3@2-1&3-11#2-5$2-5!1-2;1-2|o/C:0+0+1/D:content_1/E:content+2@2+6&2+5#1+1/F:content_1/G:0_0/H:13=7^1=2|NONE/I:13=6/J:26+13-2) and especially the starting time and the ending time. By default, the step is configured at 5ms so each segment is at least this duration.

If the labels are fine, you may check the number of values in the F0 files and if they are consistent with the number of frames you should obtained. If it's the case, reproduce the same analysis for the sp, ap, bap, mgc and cmp files.

If the counters are consistent, plot the F0 and the other components to see if they are consistent too.

Kind regards,
Sébastien Le Maguer

Martin Sulír <martin.sulir@xxxxxxxxx> writes:

> Hello,
>
> I'm trying to train new HMM-based voice in Slovak language with STRAIGHT
> algoritm, but I have some problems.
>
> Namely a lot of WARNINGs in step "Start embedded reestimation (clustered)":
>
> No path found in beta pass
> WARNING [-7324] StepBack: Bad data or over pruning
>
> and then many of these:
>
> WARNING [-2331] UpdateModels: l^k-o+u=a@2_2/A:1_1_3/B:0-0-3@2-1&3-11#2-5$2-5!
> 1-2;1-2|o/C:0+0+1/D:content_1/E:content+2@2+6&2+5#1+1/F:content_1/G:0_0/H:13=7^
> 1=2|NONE/I:13=6/J:26+13-2[25635] copied: only 0 egs
>
> and also in step: "Start forced alignment for no-silent GV", where the
> algorithm goes over 5000 beam and I have no output.
>
> WARNING [-9999] HSMMAlign: No tokens survived to final node of network at beam
> 5000
>
>
> When I use the same raw, utts and qst files with the "normal demo - Speaker
> dependent training" everithing is fine, and I get new voices without problem.
> I use 4526 raw files with 48 kHz sampling rate. I checked everithing, every
> single step in algoritm and everithing seems to be fine. When I use another
> speech database with 300 raws, I get STRAIGHT voice withou problem.
>
> I think that there might be something wrong with parameters extraction step,
> but I don't now where.
> Please, do you know where the problem might be?
>
> Thanks for your reply.
> Martin Sulír.