[hts-users:04276] Re: Label file information

[Simon King <simon.king@ed.ac.uk>]

On 20 Jun 2015, at 07:54, payman shaykhmehdi <payman.shaykhmehdi@gmail.com> wrote:

> Hi all,
> 
> Which information in the labels are optional and which of them are required?


It depends on the stream (spectral envelope, F0, etc). We found that you can reduce the number of features dramatically, with only a modest impact on quality (note: quality *is* reduced, but not a huge amount).

Lu, H., & King, S. (2012). Using Bayesian Networks to find relevant context features for HMM-based speech synthesis. In INTERSPEECH 2012, 13th Annual Conference of the International Speech Communication Association, Portland, Oregon, USA, September 9-13, 2012. 

open access version:

http://www.research.ed.ac.uk/portal/en/publications/using-bayesian-networks-to-find-relevant-context-features-for-hmmbased-speech-synthesis(c261a31e-d246-4b74-ab37-0885740e92bc).html

> for example can we skip the B section(setting b1,b2,... to zero) without significantly decreasing the quality of synthesized speech?

Which features should be retained depends on the stream, of course. For example, we found that "name of vowel of current syllable" is an important feature for spectral envelope and F0 streams, and for duration.

regards,
Simon


-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.