Dear Adam,
I am so sorry that I did not explain the problem clearly.

Actually, we found that MLP has the much larger dynamic range than PLP
has (In the log domain). Our decoder cannot deal with such large
dynamic range when decoding. I don't know if you use any method to
reduce MLP's probability range in order to adapt to PLP's probability
range.
Thank you very much!

Qingqing

--- In icsi-speech-tools@yahoogroups.com, "Adam Janin" <janin@...> wrote:
>
> --- In icsi-speech-tools@yahoogroups.com, "qingqing.zhang"
> <qingqing.zhang@> wrote:
> >
> > When I use MLP as my HMM features, the POSITIVE posteriori
> > probabilities always occurred. The POSITIVE posteriori probabilities
> > will break the decoding rule and result in bad performances. I don't
> > know whether you found the same problem, and what can I do to avoid
> > this? Thank you very much!
> >
>
> I'm not sure I understand. If you're using softmax as the output layer
> of the MLP, then all the values should be between 0.0 and 1.0. If
> you're not using softmax, the values can be pretty much anything.
>
> Perhaps your decoder wants log probabilities? Taking -24*log(p) and
> then clipping between 0 and 255 is the usual way we do this (called
> "lna format"). Also, sometimes the systems take scaled likelihoods
> rather than posteriors as input. In this case, you can just divide by
> the priors of the phones.
>
> If this doesn't answer your question, could you please post more
details?
>
> Adam
>