The result of training a neural net is a weights file which can then be used as the acoustic model to convert the features from unseen utterances into posterior probabilities of each phone class at each time step, which can then be decoded into word sequences via a decoder.
The most popular program to perform this so-called forward pass of the neural net, is qnsfwd, the forward-pass-only counterpart of the QuickNet training program qnstrn. qnsfwd takes many of the same options as qnstrn, so any on-the-fly feature processing (deltas, normalization etc.) can be directly replicated. qnsfwd also supports fixed-point calculation on the accelerated SPERT boards.
However, sometimes qnsfwd is more than is required. In particular, it requires the entire QuickNet library to be compiled, which can be difficult, and sometimes impossible (e.g. for cut-down DSP systems that may lack a C++ compiler). For these situations, we developed an alternative forward-pass program called ffwd (for 'fast forward') which is a heavily pared-down version, written in pure C, and requiring no additional libraries. ffwd reads a subset of the file formats available in the QuickNet program, and at present does not support online normalization (or delta calculation). However, it is an adequate substitute in many recognition situations. QuickNet is actually fairly optimized, but ffwd runs perhaps a little faster.
Previous: 6.4 How long and how many epochs does it take to train a net? - Next: 6.6 Tell me about the SPERT boards.
Back to ICSI Speech FAQ index
Generated by build-faq-index on Tue Mar 24 16:18:16 PDT 2009