[hts-users:02030] Re: HTS Training Error

[Geoffrey Wilfart <geoffrey.wilfart@xxxxxxxxx>]

Well, actually, as frequencies, LSP coefficients need to be in range [0-Pi] + in increasing order.

I have never observed any problem with the increasing order, but quite often seen that some coefficients get over Pi (I've never seen going under 0).

As HTS is using Gaussian distributions, there's no hard limit on a higher value of the generated coefficients.

One solution would be to constrain further the parameter generation (MLPG) to force the coefficients to stay in the range, but I haven't derived the equations. 

At this stage, I've only seen bad coefficients for a frame here and there, so another solution (not very clean, I admit) is to delete the "bad" frames.

Hope that helps,

Geoffrey

2009/6/18 wangxiang <wangxiang68@xxxxxxxxx>

Yes, my coefficients is exactly MGC-LSP.

Can you tell me more detail about the problem?

2009/6/18, Geoffrey Wilfart <geoffrey.wilfart@xxxxxxxxx>:

> Hello,
> It most likely comes from the vocoding stage, depending on the type of
> coefficients you trained.
> I first observed this on some sentences when using MGC-LSP coefficients.
> Basically,I'd think you've come up with an unstable filter.
> Can you please confirm the type of coefficients you've used ?
> Regards,
> Geoffrey
> 2009/6/18 wangxiang <wangxiang68@xxxxxxxxx>
>
>> The log file seems right, and the training process is done.
>>
>> But the synthesized wave is just slience.  And the sp file bu SPTK in
>> the synthesis stage has the matrix format:
>>
>> NaN     NaN     NaN
>> NaN     NaN     NaN
>> NaN     NaN     NaN ..
>> .....
>>
>> Which has the obvious number overflow?
>>
>> Can anyone tell me where to locate the problem?
>>
>> Thanks
>>
>>
>>
>> --
>> Xiang Wang
>> Institue Of Acoustics, CAS
>>
>>
>

--

Xiang Wang
Institue Of Acoustics, CAS