Train while correcting rec. transcription? Options

[chadlock]

After transcribing from DS3000 to DragonPad, do corrections have any influence on user files. I don't see how I can follow the "train" option in the correction box that I'm accustomed to when using my mike into the computer. Words like "I'm" always seem to come out wrong and differently.
[Background: I've been reasonably successful using Dragon Prof Sol 5, and now 6.1 at my computer. Had failed with earlier use of recorder, but just got Olympus DS3000. Did 4 training sessions, but still not good, and I know there is lots I plan to read on this forum to improve that.]

[G_Tom S_*]

Hi there.

I have the Olympus DM 1 which is almost identical to your recorder. I've done quite a bit of experimentation with its various features, and might be of some help to you.

I didn't like the recorder-users grayed-out "Training" feature either, so I ended up comparing the difference between the OPTIONS.INI files between microphone-users and recorder-users, and resetting it so that I can use a mike to do the training with the recorder's user file. They didn't make that an option with recorders, because its optimum for corrections and training to be done with the same identical mike that's used for recording.

In your case, it's my understanding that you can also use the DS 3000 as a USB microphone whereas my DM 1 doesn't have that feature. That's good for you, because with some tweaking you can set up the DS 3000 mike to use for the training and thus maintain the audio integrity of the user file.

In my case I've done some intermittent training of the DM 1 recorder user-file with my Plantronics microphone (since I can't use the DM 1 as a USB mike), but the Plantronics sound quality is so near to perfect that it doesn't match the lo-fi sound of the same words as done with the recorder, so the training fidelity doesn't quite track.

Still, I've been regularly getting approximately 95 percent accuracy with the recorder, which isn't bad considering that its main usage is intended to be out in the noisy "real world."

Keep in mind that these recorders are very excellent, but because of their sound-compression manipulations during recording they're still not engineered for creating high-fidelity audio, so you have to remove contractions (I'm = "I am", I'd = "I would", etc.) from your speech more than you're accustomed to doing with Dragon on a desktop. You also need to train yourself to separate every word, and rapidly articulate each word -- keeping tiny spaces between them -- almost to the point of being fanatic about it. The more I sound like a pedantic verbal-machine-gun with my words, the more accuracy my recorder produces in return. As a side benefit, using the recorder for awhile most definitely improves my accuracy when reverting back to desktop VR.

Krish, our Forum Administrator, told me that with a recorder user file you just need to keep patiently using the "correct that" feature and type in what it SHOULD have recognized, and Dragon will still train itself over time, even if you don't verbally train each instance.

I'd suggest that you repeatedly use "Select All" and "Play that Back" with various recordings VR'd on the screen, to spot just where Dragon diverges from your spoken word, and then take a close look at how you can maybe correct your dictation so that Dragon has an easier time of it.

If need be, it's possible to troubleshoot the hardware side of your recordings by sending me a short audio recording in Olympus' DSS format and I'll take a look at various audio traits (overall volume, dynamic volume-range, frequency response, and background noise) with Cool Edit. This would at least ensure that a good recording is making it is far as the desktop. If you do so, make sure that there's approximately 2-3 seconds somewhere in the recording where there's no talking, so I can measure the resident background noise level.

Regards,

Tom Swanson
tswans@gci.net

PS: The above message was dictated into Dragon NaturallySpeaking with a Plantronics DSP-300 noise-canceling USB microphone plugged into an AMD Athlon XP desktop computer. Message formulation, dictation, and correction took approximately 45 minutes. There were only 13 VR errors found, before correction. The length of the above text is just over 600 words.

WHERE THE RUBBER MEETS THE ROAD:

As an experiment, I decided to read the above text into my DM-1 recorder (600+ words in 6 minutes, 4 seconds = 100 words per minute), change user files, VR it against one of my DM1 user files, and then post the un-corrected results here to show you the results that my recorder produces with this particular text. Some of the errors cannot be avoided, some words/phrases simply aren't trained into that particular user file (isn't "Krisch" really our "foreign administrator"??), and some are just due to my errors in diction and pronunciation. However, keep in mind that the entire text CAN attain a 2-step 100% accuracy when correcting with Dragon's "Play That Back" (into your headphone) feature. There's no guessing what was said.


========================================


High bear.

I have the Olympus DM 1 which is almost identical to your recorder. I've done quite a bit of experimentation with its various features, and might be of some help you.

I didn't like the recorder-users great-out "Training" feature either, so I ended up comparing the differences of the options INI files between microphone-users and recorder-users, and resetting it so that I cannot user might to do training with the recorders user file. They didn't make not an option with the recorders, because it's optimum for corrections and training to be done with the same identical might that made the recording.

In your case, it's my understanding that you can also use the BS 3000 as a USB microphone whereas my DM 1 doesn't have the feature. That's good for you, because with some tweaking you can set up the BS 3000 might use for the training and us maintain the audio integrity of the user file.

In my case that doesn't intimate and training with the DM 1 recorder user-file with the Plantronics microphone (since I can't use the DM 1 as a USB might), but the Plantronics sound quality of sundered imperfect that it doesn't match the Lowell-by sound of the same words is done with the recorder, so the training fidelity doesn't quite track.

No, I been regularly getting approximately 95 percent accuracy with the recorder, which isn't bad considering that its main usage is intended to be out in the noisy "real world."

Keep in mind that these recorders are very excellent, the because of their sound-compression manipulations during recording is still not engineered for creating high-fidelity audio, so you have to remove contractions (I'm = "I am", I'd = "I would", etc.) from your speech more than you're accustomed to doing what Dragon on the desktop. You also need to train yourself to separate every word, and rapidly articulate each word -- keeping tiny spaces between them -- almost to the point of being fanatic about it. The more I sound like it pedantic verbal-machine-gun with my words, the more accuracy my recorder produces in return. As a side benefit, using the recorder for a while most definitely improved by accuracy undergirding back to desktop VR.

Krisch, our foreign administrator, told me that with the recorder user file you need to keep patiently using the "correct that" feature and I been leaded SHOULD have recognized, and Dragon all still trained itself even over time, if you don't verbally train each instance.

I'd suggest that you repeatedly use "Select All" and "Play that Back" with various recording VR' Chris B. on the screen, dewpoint does were Dragon diverges from your spoken word, and then take a close look at how you can maybe correct your dictation so that Dragon has the an easier time of the area

If need be, it's possible to troubleshoot the hardware side of your recordings by sending me a short audio recording and Olympus' the S. S. format and I'll take a look at various audio traits (overall volume, dynamic volume-range, frequency response, and background noise) with cool edit. This would at least ensure that a good recording is making it as far as the desktop. If you do so, make sure that there's approximately 2-3 seconds somewhere in the recording were there's no talking, so I can measure the resident background noise level.

Regards,

Tom Swanson

[Krish]

>>words/phrases simply aren't trained into that particular user file (isn't "Krisch" really our "foreign administrator"??)>>

haha...That is Great!!

excellent post, thank you for the stats and details.

[G_Tom S_*]

You're welcome, Krish. This is fun.

--As our Foreign Administrator, just how many foreign languages do yawl need to speak with, anyhow? (North Texas, South Texas, East Texas, West Texas, Floridian, and just whatever else we'uns can cipher out?

Tom

[chadlock]

Tom -- Thank you for your detailed reply. I did some more experimentation over the weekend and appreciate your offer to check out the parameters for one of my dss files, which I will E-mail to you separately unless I find a way to attach it to this post. There are two paragraphs there, and my speaking was a bit more subdued in the second, but the overall transcription accuracy seemed comparable. Let me ask about two other things mentioned in your message.

First, I don't see any way to use the DS3000 as a USB mike, but there are some vague comments in the users manual about dealers being able to configure some things. Do you know if I can use it this way. (I got it used off Ebay and have no "dealer".)

Second, where do you and others get the accuracy stats (e.g. 95%) so often quoted? Are these just estimates or can the program track corrections to calculate this?

Thanks again for your advice.

Charlie



[This message has been edited by chadlock (edited 02-03-2003).]

[G_Tom S_*]

Hi Charlie,

I once saw a reference saying that the DSS Player Pro software is the key to using the 3000 as a USB mike. Did you get that program with your recorder? I hope it's not an extra option that doesn't come with the recorder. I saw the claim that the recorder can be used for a USB microphone posted on about a dozen websites, so I believed them.

If it's not an optional item, tell Olympus your story and maybe they'll get you a copy of the software.

Check out this webpage for a bit of interesting DS3000 information/gossip: http://www.voicerecognition.net/archive/ht...9/msg00000.html

You can't attach files to posts in this forum, so just email it to me direct, at tswans@gci.net. I'm standardizing a system of developing public feedback for audio files that I've plugged into CoolEdit to be analyzed for sound qualities and deficiencies, and this feature should get more interesting as time goes on.

Incidentally, are you considering connecting a headset mike to the DS3000? Since there's a noticeable difference between the sound-level of the audio that's output by various mikes, and since the DM1/DS3000's mike input circuitry is a bit more sensitive than the standard MIC IN jack on a soundcard, then the outputs of standard external mikes usually need to be cranked down just a bit. I found a variable mike volume control that seems to work pretty good for that. Radio Shack sells a tiny, inexpensive in-line volume control intended for speaker/earphones, but it also works great as a mike attenuator for our particular purposes. Just keep in mind that any mike changes could mean training a new user file for that mike, unless it is VERY similar in output (sound quality and volume level) as another one.

I too had hopes that DNS automatically calculated the percentage of VR accuracy, but it doesn't. There are probably too many variables for a machine to do so, because some corrections are probably dependent upon your personal opinion and choice as to what you REALLY wanted it to type out.

Keep this in mind:

1 error per each 100 words leaves 99 words correct, or 99% accuracy.
2 errors per each 100 words (one in 50) leaves 98 words correct, or 98% accuracy.
5 errors per each 100 words (one in 20) leaves 95 words correct, or 95% accuracy.
etc.

To calculate VR error rate, I first count how many overall errors I corrected, and then get a total word-count by temporarily pasting the finished, corrected document into MS Word and selecting TOOLS - WORD COUNT.

After that, it's simple math. A workable formula is 1 - (1 / (total words / total errors)).

For example, assume there were 70 errors in a 3,300 word document. At first glance, simple division reveals this to be 47 accurate words per each single error, or 1/47 error rate, which is just a bit worse than 1 error per each 50 words, or 98% accuracy.

Plugging the numbers into the formula, 1 - (1 / (3,300 / 70)) comes out to .978788, or approx. 97.9% accurate. (Thanks, MS Excel!) :-)

More soon,

Tom Swanson



[This message has been edited by Tom S (edited 02-04-2003).]

[John Wickett]

Tom,

I presume that when you are measuring accuracy you are not counting errors on your part as DNS errors. If you say "people" and meant to say "persons" then that is not an error to be included in the mix.

When Judy, Ross, myself and others are reporting accuracy rates we are not counting misspoken words unless DNS gets them wrong. Just want to make sure that you are using the same method we are.

John

[CB HINSON]

HEY BOY!Yawl air in aheepa of trouble trying to talks Southern.Yawl got to be raised hern to do that. But bless your cotton picking heart far trying.

CB

[G_Tom S_*]

John,

You presume correct. It's common for me to proofread something and repeatedly change the wording until it finally sounds right, which I try to keep separate from the actual VR step that Dragon initially performed. When calculating Dragon's VR accuracy I only calculate the errors that Dragon made in deciphering the sounds of my voice, or else the accuracy rate would be in the area of 30% or so, if I foolishly included all of my "creative" changes on top of that!!!! :-)

Thanks,

Tom Swanson

PS: CB, you're a hoot. No, not a Hooty-Owl, just a plain ol' HOOT. Hot-dang, I'd like to see your VR results.