[hts-users:03371] Re: question about GV Based parameter generation

[soheil khorram <soheil_khorram@xxxxxxxxx>]

Thanks for your answer,

But I guess I have to explain my question.

The algorithm consider that the GV of each parameters is time invariant and is identical for each contexts. But when we train a decision tree for GV distribution, it causes a context dependent distribution and the distribution may be changed in time.

---

From: Keiichiro Oura <uratec@xxxxxxxxxxxxxxx>
To: hts-users <hts-users@xxxxxxxxxxxxxxx>
Cc: uratec <uratec@xxxxxxxxxxxx>
Sent: Tuesday, June 26, 2012 12:20 AM
Subject: [hts-users:03370] Re: question about GV Based parameter generation

Hi,

The algorithm is not changed.
Only model tying structure is changed.

Keiichiro Oura, Yi-Jian Wu, and Keiichi Tokuda,
Overview of NIT HMM-based speech synthesis system for Blizzard Challenge 2009,
Blizzard Challenge 2009, 2009-09.

Regards,
Keiichiro Oura


2012/6/26 soheil khorram <soheil_khorram@xxxxxxxxx>:
> Dear all,
>
> I have read the paper titled "A Speech Parameter Generation Algorithm
> Considering Global Variance for HMM-Based Speech Synthesis" and I am
> confused about the GV distribution. According to the paper, the GV
> distribution for each parameter is a single Gaussian, but in the HTS demo a
> decision tree is trained for this parameter.
> I want to know, when a context dependent distribution is used for GV, what
> part of the algorithm will be changed.
> I will be appreciated if there is a useful reference for it.
> Thanks in advance.
>
> Khorram