[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:03673] Re: help: mel cepstral distortion measure


Hi,

I am also doing the same stuff recently. I am trying to analyse and re-synthesise the speech without statistical modelling. As different analysing ways use different spectrum parameter to re-build speech, if I use cdist in SPTK to judge which analysing way is better, the objective way is quite different from the subjective result. So can I ask are there any other better ways to judge the system from objective method (still without HMM building) ?

For cdist to calculate the distance of cpestrum, is it necessary to get the cepstrum from original spectrum parameter it uses, or I could still use reconstructed wave to get the cepstrum and then calculate the distance?

Thanks.

Ellis


On 7 Mar 2013, at 16:37, Keiichiro Oura <uratec@xxxxxxxxxxxxxxx> wrote:

> Hi,
> 
> I think that the dynamic time warping (``dtw command in SPTK) is one
> of the good solution.
> 
> Another solution is speech synthesis by using forced-alignment label
> of natural speech.
> HSMM with -f option and HMGenS with -s option can do it.
> 
> Regards,
> Keiichiro Oura
> 
> 
> 2013/3/7 Baji Babu <bajibabu7@xxxxxxxxx>:
>> Hi,
>> 
>> What people normally do when two sentences are not aligned is that they use
>> the dynamic time wrapping (DTW) technique to normalize the durations.
>> 
>> Regards,
>> Bajibabu B
>> 
>> 
>> On Thu, Mar 7, 2013 at 3:34 PM, Hardik Sailor <s.hardik89@xxxxxxxxx> wrote:
>>> 
>>> But first I want to synthesize speech of same duration like natural one. I
>>> don't know how to do this. Then I can use cdist and rmse commands of SPTK.
>>> Please sir guide me on this. I am using HTS demo.
>>> 
>>> Thank you Sir
>>> 
>>> 
>>> On Thu, Mar 7, 2013 at 7:00 PM, Keiichiro Oura <uratec@xxxxxxxxxxxxxxx>
>>> wrote:
>>>> 
>>>> Hi,
>>>> 
>>>> Could you try to use ``cdist'' or ``rmse'' commands of SPTK?
>>>> 
>>>> Regards,
>>>> Keiichiro Oura
>>>> 
>>>> 
>>>> 2013/3/5 Hardik Sailor <s.hardik89@xxxxxxxxx>:
>>>>> Hello,
>>>>> 
>>>>> I am new to HTS World. I have demo version of HTS with CMU arctic SLT
>>>>> voice.
>>>>> As a part of my research work,
>>>>> I want to find quality and intelligibility and other measures for HTS.
>>>>> So I
>>>>> want to know how to find MCD which require same lengths of natural and
>>>>> synthetic speech. And also RMSE measure. I am right now, working with
>>>>> HTS
>>>>> SLT demo.
>>>>> 
>>>>> Thank you
>>>> 
>>> 
>> 
> 


References
[hts-users:03650] help: mel cepstral distortion measure, Hardik Sailor
[hts-users:03656] Re: help: mel cepstral distortion measure, Keiichiro Oura
[hts-users:03657] Re: help: mel cepstral distortion measure, Hardik Sailor
[hts-users:03658] Re: help: mel cepstral distortion measure, Baji Babu
[hts-users:03659] Re: help: mel cepstral distortion measure, Keiichiro Oura