Xref: pavo.csi.cam.ac.uk comp.speech:713 comp.ai.neural-nets:3092 Path: pavo.csi.cam.ac.uk!doc.ic.ac.uk!marble.uknet.ac.uk!uknet!cam-eng!jam!ajr From: ajr@jam.eng.cam.ac.uk (Tony Robinson) Newsgroups: comp.speech,comp.ai.neural-nets Subject: ANNOUNCE: recnet-1.0: A recurrent neural net speech recognition system Message-ID: Date: 5 Apr 93 22:24:47 GMT Sender: ajr@eng.cam.ac.uk (Tony Robinson) Organization: Engineering Department, Cambridge University, England. Lines: 63 Nntp-Posting-Host: jam.eng.cam.ac.uk The package recnet-1.0.tar is now available from the "anonymous" ftp site svr-ftp.eng.cam.ac.uk in directory misc. recnet is a complete speech recognition system for the DARPA TIMIT and Resource Management tasks. It uses recurrent networks to estimate phone probabilities and Markov models to find the most probable sequence of phones or words. The system is a snapshot of evolving research code. There is no documentation other than published research papers. It is configured for the two specific databases and is unlikely to be of use as a complete system for other tasks. This is my own research code and unfortunately I have little time to support its use by others (unless paid for in terms of joint publications of course). The components are: A preprocessor which implements many standard and many non-standard front end processing A recurrent net recogniser and parameter files Two Markov model based recognisers, one for phone recognition and one for word recognition A dynamic programming scoring package The complete system performs competatively. I believe the TIMIT phone recognition rates are the best reported (30.7% errors for 61 symbols). The RM system is better than any other monophone based recogniser (about 5% error on Feb89, Oct89 and Sep92 test sets), although not yet as good as the best triphone based systems (3-4% error on same test sets). Publications giving details of these results can be found on svr-ftp.eng.cam.ac.uk in directory reports (consult the ABSTRACT file first). I hope this will be of use to a wide audience, for example: Researchers who are familiar with TIMIT and RM and who want to compare their systems to another Researchers who need specific chunks of code (e.g. LPC or DP scoring) and who are prepared to hack it out Researchers who want some intermediate results such as phone confusions or an initial phone alignment for Resource Management Researchers who want to construct a similar system of their own To make life a little easier, I've set up a Makefile so that "make TIMIT" or "make RM" does everything (tested on SunOS and IRIX). You will need the relevant CDROMs, 150MByte of free space for TIMIT and 300MByte for RM. The file recnet-1.0.tar totals 2Mbyte in size including 7,000 lines of uncommented C source, so you have to be pretty serious to make use of this offer. If you do use the code, I'd appreciate an email note so that I can keep you informed of new releases. Tony Robinson ------------------------------------------------------------------------------- Tony Robinson Phone +44 223332815 Fax +44 223332662 Email ajr@eng.cam.ac.uk Cambridge University Engineering Department, Trumpington Street, Cambridge, UK. -------------------------------------------------------------------------------