QUOTE (bluerhino @ Apr 16 2009, 07:18 PM)
Hello everyone. I need to use voice recognition but have not been able to figure out how to use it sucessfully.
Dragon Naturally Speaking 10 Question:
One of my problems with Dragon seem to be based on an exponential deterioration of voice recognition. The program seems to recognize phrases and commands initially, but as time progresses the program seems to perform worse and worse. Some days it runs poorly from the jump, and others it is better. The program often seems to believe that phrases or words I say are commands, and sometimes it even refrains from responding to the most obvious commands that I give it. For example, When correcting a word you have the option of one of four corrections. I will say "Choose 2", and the program simply does not respond, or it will move and choose another choice altogether. I have increased my ram to 4 GB of DDR2 running at 800mhz, defragged my HDD, ran the program through several training programs, added new words to its recognition list, and I still feel that it is not learning anything that I teach it. Has anyone here had similiar problems? Is there anything that I am personally doing? Is my computer powerful enough to run this program efficiently?
Dell Dimension C521
AMD Athalon 64 Dual-Core Processor 4000+ (2.10 Ghz)
4GB DDR2, 800 mhz
Windows Vista 32-bit
Core 2 Duo (1.5Ghz)
4.00 GB DDR2 (800mhz)
Windows Vista 32-bit
Unfortunately, the one thing that would be helpful in determining why you are not getting good accuracy with DNS 10 is omitted from your post. That is, what microphone/soundcard are you using? Make sure you let us know what your make/model of microphone is as well as the same for your soundcard.
When you run the Audio Setup Wizard volume adjustment where does the blue bar fall (end up) during the microphone volume adjustment phase? Above 75%, below 25%, approximately 50%, what?
When you run the Audio Setup Wizard Quality Check, how high are the yellow bars in the graph why you are dictating the text? How much space is there between the top of the green bars and the top of the yellow bars? To the green bars extend almost to the top of the display area? If not, where do they fall?
When you run the Audio Setup Wizard Quality Check what is the Quality Check Ratio number that you get at the end?
When you complete the Audio Setup Wizard and before exiting the wizard, click the Play button. What is the quality of your voice reply. Is it clear and easily distinguishable (remember that the human ear is 100 times better than the computer at discriminating audio)? Is it muffled? Is there any background noise (i.e., low-frequency hum, crackling or what is normally called white noise [hiss])? Can you hear other background noise (i.e., radio, other people speaking, room sounds, etc.)?
When running the Audio Setup Wizard, Quality Check test portion, jiggle the wire on your microphone (if you are using a wired close talk microphone) and see if it produces any spikes are introduces any crackling noise. Also, do this a couple of times paying particular attention to the connection points to the input Jack and the connection to any mute switch or to the headphones. It is possible that you have a defective microphone, though unlikely, which is shorting out due to broken wires or connections.
Reason that I ask these questions is that the most probable cause for your problem has nothing to do with DNS itself, but has to do with the quality of your microphone/soundcard. That piece of hardware is significantly more critical to accurate speech recognition than any other part of your computer system. Also, if your dictation style, enunciation, pace of speaking, etc. are not natural, this has an impact. Overtraining. That is, you should never have to train more than once, after which you should get at least 98% accuracy. In fact, if you select "None" as the initial general training option when creating a new user, that is the speaker independent Acoustic Model and most users get at least 97% to 98% without ever doing any training whatsoever on that user profile selection. Regardless, if you have to train your user profile more than once, or if you do training more than three times, and you still get poor accuracy, then your Acoustic Model is not optimal. Running multiple training sessions is not recommended because on a poor initial Acoustic Model, multiple trainings tend to degrade user accuracy anyway.
Keep in mind that your experience relative to the accuracy that you are obtaining from DNS 10 is the extreme exception rather than the rule. Therefore, there is a reason for this and it is most likely in your microphone/soundcard or the manner in which you dictate, or both.
I await your responses to my questions, after which you will take it to the next level.
Owner, GEMCCON - The Choice of Intelligence
Speech Recognition Consulting and Training
If there's more than one way to do a job, and one of those ways will result in disaster, then somebody will do it that way. (variant of Murphy's law - Edward A. Murphy, Jr.)