-- Main.matthewa - 27 Jul 2006

At present I've implements the Needleman-Walsh algorithm for aligning letters to phones in the LTS system I've built. This is basically the same as a Levenstein style match. Viterbi training maximises the probability of this search. However this is based purely on the P(phone, letter). This doesn't involve any structure. That is substructure like 'th' etc.

We need to discuss this as I have the feeling the alignment might be crucial in how well the system performs in the end. I've had a search for sequence matching algorithms and emailed Miles Osborne to see if what I'm doing is sensible.

Suggest you also talk to other MT people in Edinburgh such as Philip Koehn or Chris Callison-Burch (both here in Baltimore just now though!) Framing the LTS problem as machine translation sounds interesting, although we don't have anything like the size of parallel corpora that they have. May still find some techniques are applicable - phrase tables and so on -- Main.simonk - 31 Jul 2006

I will get in touch with them. I got as far as writing a little script to find n-grams in a lexicon and compare their actual frequencies with expected frequencies given the frequency of the n-gram (n - 1) and the frequency of the next symbol. This should give some idea of what sequences you may want to align to other than unigrams (in both ephones and letters). I'll get back to this when I've had a chance to talk to the MT people. -- Main.matthewa -3 Aug 2006

Topic revision: r3 - 03 Aug 2006 - 09:50:44 - Main.matthewa
This site is powered by the TWiki collaboration platformCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback
This Wiki uses Cookies