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

[hts-users:01453] Re: MCP vs LSP


Hi,

Nicholas Volk wrote (2008/06/23 18:04):

I see that in the 2.1 RC2 SLT demo the hts_engine uses MCP.
Is there some particular reason, why beta is set to 0.0 in it?

Because GV is used in the 2.1 SLT demo.
If we set beta>0.0, postfiltering is turned on.
However, GV works similar to postfiltering and better than postfiltering.
Therefore, we turned off postfiltering in the 2.1 SLT demo.
Cascading GV and postfiltering is possible, but I think it's nonsense.

Are there any listening tests that would tell us which is better, LSP or MCP?

It depends on the case, but Mel-LSP often works better than mel-cepstrum for speaker dependent system. Furthermore, Kim et al. reported that 18-th order LSPs achieved almost the same quality as 24-th order mel-cepstral coefficients in S.-J. Kim, J.-J. Kim, M.-S. Hahn, HMM-based Korean speech synthesis system for hand-held devices,
IEEE Trans. Consumer Electronics, vol.52, no.4, pp.1384-1390, Nov. 2006.

It should also be noted that the synthesis filter for LSP-type parameters usually requires lower computational cost than that for cepstral ones.

For speaker adaptation system, it is difficult to guarantee the filter stability of generated parameters if LSP-type parameters is used.

What gamma value should be used for LSP?

It also depends on the case, but as gamma increases (gamma=-1,-1/2,-1/3,...) the computational cost increases.

Regards,

Heiga ZEN (Byung Ha CHUN)

--
------------------------------------------------
Heiga ZEN     (in Japanese pronunciation)
Byung Ha CHUN (in Korean pronunciation)

Department of Computer Science and Engineering
Nagoya Institute of Technology
Gokiso-cho, Showa-ku, Nagoya 466-8555 Japan

http://www.sp.nitech.ac.jp/~zen
------------------------------------------------

References
[hts-users:01452] MCP vs LSP, Nicholas Volk