On 16/02/2008 at 00:07 mal wrote:

>Take the letter Z for instance sent badly spaced. It could be TD, MI, GE, 
>TTEE. A proficient cw operator would be able to identify the context in which 
>it was being sent and read it accordingly. Presently it could not be >read by 
>a machine.

If you were doing machine based CW recognition for serious, you would start by 
collecting data on how people mangle code.  This is similar, though with a much 
more limited language, to the data collection work that was done in order to 
produce Speaker-Independent recognition of spoken digits for the telephone 
service. 

I expect you would find a number of common errors (equivalent to different 
accents in spoken language), as for example a contester who calls NST when he 
means TEST, and of course 5NN for a signal report*.  Then there are the bug key 
users who speed up their sending by taking the weights off to shorten the dots 
but do not make the corresponding adjustment in the manually-produced dashes.  
Once you have a sufficient quantity of data, you would be in a position to 
analyse it and establish how many different "accent" you would have to deal 
with.

Do not expect to produce a system that is 100% accurate.  The ear plus brain 
doesn't provide that.

John F5VLF

*does anyone use short numerals nowadays?