DNS 9.5 Professional user files
Thread poster: Transilvania
I have used DNS for a while but I have a Problem:
When I create a new user and train the program and read the sample texts everything works fine after that. Than I used to read a few common economy texts from a magazin fluently and clearly and saved the user data. Up to this point the software works good not really perfectly when it comes to recognition but it's ok.
But when I start to work and train the program with my own vocabulary and safe the results I cannot translate so fluently and the software makes more mistakes because I think it is confused due to the corrections I have made before and it makes wrong conclusions for other words. I have used the correct approaches to correct the texts with DNS options and did not type any corrections with the keyboard . But after a while the recognition becomes worse and worse and the vocabulary, especially the letters at the end of words like plural and singular are misunderstood. Also the software seems to be confused which words it should capitalise and which not or write in one word or seperatley (we are talking about German language here). So I can only use the software if I don't save the user files after work, otherwise the recognition is worse than before.
I think DNS is not up to date with the latest changes in the German language. Nowadays you have so many different opportunities to write a word as a result of the language reform and the reform of the reform. It seperates words which would be perfectly ok to write as one word now and which I just cannot accept to write seperatly because it looks strange to me and to my customers. For my version of DNS the old language files prior to the reform are not compatible.
Is anybody confronted with the same problem?
Can I somehow safe my user files after work with another user name that I don't destroy my original user file but don't always have to use the exported backup files? I mean...can I create a new user from the basis of an existing user?
| || || |
| a few tips... || Oct 25, 2007 |
I can't really give you an answer to your question, Transilvania, only a few tips here and there.
DNS performs best when you dictate without interruptions. Once you start dictating individual words, the recognition factor deteriorates rapidly. That's the algorithms on which the two speech recognition applications (DNS and ViaVoice) are based. After all, DNS was designed with users in mind who dictate whole documents in one go, whereas we translators tend to correct a word here and a phrase there. DNS is not partuclar good at that (neither is ViaVoice) but I'm afraid we have to live with it.
Whether you add corrections through your keyboard (which I frequently do for the sake of speed) or via microphone is IMHO not all that important, provided you do the keyboard correction in the appropriate box.
Capitalization in German is a pain in the neck, I admit, and frequently DNS is not in a position to determine whether e.g. "Ihnen" oder "ihnen" should be used. You can work around this problem by dictating "groß ihnen" (or "groß [any word]"), and DNS will capitalize the word following "groß"; in English, the corresponding voice command would be "cap" or "caps", I believe.
If you don't like the spelling version DNS presents you with, you can actually correct it in the user area so DNS will do it right ever after. I think you may want to read up on this in the manual; discussing this area here would be beyond the scope of this forum.
I don't quite follow your reasoning as to why you have multiple users. I have two users, one for German>English, one for English>German, as I need two distinct users for this combination. The longer the software gets used to me as the same user, the better the recognition, provided I train it right. Any previous training of the software is lost, AFAIK, if I open a new user (i.e., that training is not available to the new user and would have to be done over again). What a waste of efforts!
Altogether I can't complain about the recognition factor of DNS (or ViaVoice for that matter). There are certain specialized areas where the recognition is low; I am aware of them and occasionally forego the use of DNS altogether if I have such a translation. In areas that I am specialized in (legal, business), though, the performance is very nice, even though I've been using DNS for little over a year only. Before that, I ran ViaVoice. Now, I use ViaVoice only as a text-to-speech application, as it is very cumbersome when used as speech recognition application in SDLX.
Keep up the good work, it will pay off eventually...
| || || |
| 2 different issues || Oct 25, 2007 |
It took me a long time to realize it, but there are 2 different issues at work here.
But when I start to work and train the program with my own vocabulary and safe the results I cannot translate so fluently and the software makes more mistakes because I think it is confused due to the corrections I have made before and it makes wrong conclusions for other words. I have used the correct approaches to correct the texts with DNS options and did not type any corrections with the keyboard .
First, make sure that you speak in full phrases and when you correct, you correct a phrase of 3-5 words or more. DNS figures out what you are saying by looking at the probability of any 2-3 words appearing together. It needs the context from you in order to make that determination correctly when you speak, and it also needs the context when you correct (your corrections don't update its pronunciation model so much as its probability tables), so try to get a few words around the one that was mistaken.
I also find in English that word endings are problematic. Final s's and past-tense d's tend to be obliterated, which may be a symptom of my own particular brand of muttering. If I am aware that a problem combination is coming up (say, "these stocks" or "referred to" which respectively tend to be transcribed "the stocks" and "refer to"), I try to make a slight pause between them, which helps improve things a bit.
Also the software seems to be confused which words it should capitalise and which not or write in one word or seperatley
This is actually a separate issue and entirely a function of the stupidity with which DNS's vocabulary was put together. If you get random capitalization or compound/non-compound words, it's probably because DNS contains both forms in its vocabulary and is just tossing a coin as to which one to use. For instance, in English both "association" and "Association" appear commonly in written speech, so DNS has both forms in its vocabulary and you are just as likely to get one or the other for any given utterance.
Since I translate into English, I prefer to have my words in lowercase unless I specifically instruct to the contrary. Germans, I realize, have different preferences for their nouns. Here's how I ensure consistency. Be warned that it's a royal pain in the a**, but it does work:
When confronted with a random capitalization, open the vocabulary editor, find the offending form of the offending word, select it and click "properties." The tab that opens up will provide a checkbox that says "use alternative written form." Check this, and then in the blank write the form you want to see as the default (in my case, the lowercase). Thereafter, it doesn't matter which form DNS thinks it has heard, you will get your default form. If you choose the lowercase form as default and happen to need it capitalized, you can still get capitalization by saying "cap" before the word; if you choose capitalized and want lowercase, just say "no caps" first.
You can do the same thing for compound/non-compound words. If you always want a compound word split up, find the compound form and use "properties" to give it the proper written form. If you always want the compound, enter the non-compound form as a vocabulary item and then give it the compound form as the "alternative written form."
Be warned, you'll be doing this several hundred times before you've got something usable, but at the end of the process it will indeed be usable.
(And all of it could have been avoided if DNS had not padded its vocabulary...)
| || || |
| | Transilvania
Local time: 09:22
English to German
| Thanks a lot both of you... || Oct 25, 2007 |
......for your detailed reply!
I will try and see what happens in the next few weeks and check back.