From your message, I assume you're using the MLP outputs as features
to a GMM (so-called "tandem") rather than using the posteriors
directly (so-called "hybrid"). In this case, we either use a linear
output layer, or use log probabilities. In both cases, we usually
append acoustic features, and follow it up with PCA or LDA to reduce
the dimensionality.

I can check with folks here, but I don't recall any normalization
steps prior to this, so I can see how dynamic range might be a
problem. You might try just doing mean/variance normalization.

Good luck,

Adam


--- In icsi-speech-tools@yahoogroups.com, "qingqing.zhang"
<qingqing.zhang@...> wrote:
>
> 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!