Fwd: asr: r00_n10: N-grams with cross-word models

[robert b <http://dummy.us.eu.org/robert>]

--- Joe Picone <http://www.isip.msstate.edu/~picone> wrote:
> Subject: asr: r00_n10: N-grams with cross-word models
> Date: Tue, 7 Jan 2003 21:26:00 -0600
> From: "Joe Picone" <http://www.isip.msstate.edu/~picone>
> To: <http://www.isip.msstate.edu/~asr>
> 
> We are pleased to announce release r00_n10 of our
> recognition tools:
> 
>  http://www.isip.msstate.edu/projects/speech/software/
> 
> Please see our monthly tutorial:
> 
>  http://www.isip.msstate.edu/projects/speech/software/tutorials/monthly/2003_01/index.html
> 
> for an overview of all the features of this release. It represents
> a substantial upgrade from r00_n09. The system is extremely flexible
> and user configurable. For example, cross-unit decoding
> (e.g., cross-word phonetic models) can be done at any level in the
> hierarchy of representations (e.g., sentence, part of speech, word,
> phone, state). We have also added support for Annotation Graphs
> and database access to audio files. There is even a stack decoder
> in this release.
> 
> The on-line tutorial:
> 
>  http://www.isip.msstate.edu/projects/speech/software/tutorials/production/fundamentals/current/
> 
> is being developed using this version of the system. You will find
> many useful step-by-step instructions how to use this system
> at this URL.
> 
> Let me take a moment to discuss a few strategic issues related
> to our project.
> 
> First, this is the last release for which we will support Sparc
> processors. Even Sun has been moving away from Sparcs, and we now
> have limited outdated cycles available on Sparcs. We can barely run
> the necessary regression tests because our Sparcs are now considered
> too slow (450 MHz) compared to our x86 cycles. In the future,
> we will only test our releases on Solaris and Linux for x86 systems.
> 
> Second, this release compiles cleanly for gcc 3.1. We will support gcc 3.2
> soon, but wanted to make this release first. Unfortunately, there are 
> significant differences between 3.0, 3.1, and 3.2, so it is hard to make 
> this code stable for new releases that we haven't seen yet.
> 
> Third, we expect one more release by May 1 that will be primarily
> an efficiency release. This will be r01_n00, and represent an
> end to this phase of our public domain project. The decoder in r00_n10 is
> about 8x slower on N-gram/cross-word decoding than our prototype system.
> Work is underway to close this gap.
> 
> Fourth, because we are having a very bad funding year, work on this
> project has slowed to a crawl. We have a skeleton staff working
> on a few things, but we can't do much more until we find additional
> sources of funding for this work. Hence, I can't promise many more
> releases beyond r01_n00, though we remain optimistic that this project
> will continue beyond May 15.
> 
> The current release represents what we believe to be the most flexible
> decoder publicly available. It has hooks for many different types of
> research. We are now actively using this system on a number of internal
> research projects, and will release some of this research soon.
> 
> Thanks to Issac Alphonso, Kaihua Huang, and Naveen Parihar for making
> this release possible.
> 
> -Joe