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

[hts-users:03859] Re: objective evaluation


 
 Dear Matt.
I think that my formula of lf0 distance change as below.
본문 이미지 1
Is right?
I want to receive confirmation from you.


2013/9/24 Hea Young Park <lovelyoe.park@xxxxxxxxx>
Thank you, Matt.
 
But I still wonder the method that calculate lf0 distance.


2013/9/23 Matt Shannon <sms46@xxxxxxxxx>
Hi,

I believe the conventional way to compute the MCD between a reference
sequence of cepstra and a generated sequence of cepstra (which are
already aligned in time in some sense) is to compute the root sum
squared error for each frame, then average across all frames, then
multiply by the constant sqrt(2.0) * 10.0 / log(10.0).  Your formula
does not have the constant or the square root after the sum over $t$.
It's also conventional to leave out mgc0 from the computation.

Matt


On 23/09/13 02:49, Hea Young Park wrote:
> Hello, everyone.
> I did objective evaluation about HTS system that I had built.
> Below is my calculation about each distance.
> 본문 이미지 1
>   Unfortunately, as you can see, the range of mgc and lf0 distance from
> my objective evaluation is so many different from that of the papers
> (see, table below)
>
> 본문 이미지 2
>
> Please let me know if you find something wrong in my calculation.
>
> Below is the result of what I calculated.
>
>   - mgc distance : 1.948
>
>   - lf0 distance : 0.133
>




References
[hts-users:03853] objective evaluation, Hea Young Park
[hts-users:03857] Re: objective evaluation, Matt Shannon
[hts-users:03858] Re: objective evaluation, Hea Young Park