More RAM when you already have 2GB won't help your accuracy or speed. Its likely you're getting lazy in your usage, which is typical when you increase your usage. Most productive for you would be to describe in as much detail as possible how you dictate, correct, etc., and let people suggest improvements in your practices.

However, I would start by suggesting you playback every error to hear what you actually said and only correct when you actually said what you thought you said. Otherwise, just re-dictate.

HTH,
Bruce

Bruce is basically correct. However, allow me to clarify a couple points even though this issue has been discussed numerous times in the past and you can find all the discussion threads if you do a search on this and other forums.

First, increasing RAM has virtually nothing to do with DNS. DNS already loads everything into RAM when you launch your profile.

Second, what increasing RAM does is increase the speed and performance of Windows. The faster and more efficiently Windows runs, the faster and more efficiently DNS will run. So the consequence of adding RAM, wherein it increases the performance and speed of Windows, indirectly "CAN" indirectly increase the performance and speed of DNS. In an of itself, however, it does not directly impact on DNS. What does improve the performance of DNS is the amount of L2 cache available, which I will explain in a moment.

Third, and I have never gone back and discussed this thoroughly in the last couple of years, the DNS recognizer (MREC.dll) does not use the CPU to perform its analysis and conversion of speech to text (dictation to transcription). That function is performed by the FPU (Floating-Point Processor or coprocessor as it used to be called many years ago). What made speech recognition possible on PCs was the incorporation of the FPU as part of the processor unit with the advent of the building the coprocessor into the CPU chip that came about with the advent of the 486 and the AMD K-6. When this happened, Microsoft introduced MMX, which automatically channeled mathematical computations through the FPU. Since recognition is entirely a mathematical process, the entire process of converting speech into a mathematical model (Acoustic Model) and comparison to the Vocabulary and Language Models is performed in the FPU, not the CPU. This is because the entire process is based on the mathematical models used in the HMM (hidden Markov models) that comprise the Acoustic Model and the Language Models.

Fourth, and this goes back to the second point, each step along the process saves its results to the level 2 cache (L2 cache) for quick access by the FPU when the next step in the process is executed and requires the results of the previous process. For example, when you dictate something (each separate utterance) the Acoustic Model compares your pronunciation to the phonetic table contained in the Vocabulary (i.e., the active Vocabulary that is loaded into RAM). It then takes this parsed set of phonetic strings and converts them to a statistical probability model based on the algorithms associated with creating an HMM equivalent. DNS then takes this converted HMM model based on your dictation and searches all of the possible HMM equivalents already encoded in the Language Models. This process is governed by the position of the Speed vs. Accuracy slider in the Miscellaneous tab of the DNS Options dialog. This is basically a timer and determines how long DNS searches through the various statistical and probability combinations of words (this is not a database of words it is a statistical model that is based on probability coefficients relative to parse strings not words). Because this is a mathematical computation and uses the FPU it is extremely fast (i.e., basically in milliseconds). DNS then selects the top 9 highest probability results from the Language Models. The result of this computation is then passed back through the L2 cache and is accessed by the CPU in order to compare it to the Vocabulary for the parsing of the binary phonetic results and conversion into words that are then passed through SAPI back to the application into which are dictating. So, basically all of the computations are done by the FPU, but the transcription process (i.e., converting the results to text) is performed by the CPU because only the CPU can execute Windows instructions, and SAPI is a set of Windows instructions specifically designed for speech. That is, passing speech information back and forth between applications whether dictation or commands.

Therefore, the end result is that the amount of RAM has very little to do with this process. It does have a great deal to do with how efficiently Windows performs. Windows consumes about one half to two thirds of your available RAM for its own purposes. If you have only 1 GB of RAM, some of the RAM has to be converted for use entirely by Windows, some of the RAM is allocated to applications that are active and loaded, some of the RAM is allocated to resources (i.e., tracking everything that is running), and some of the RAM is allocated to those resources necessary for keeping track of user and graphical user interfaces. What you have left over is then accessible to applications. In the case of 1 GB of RAM, you can end up with insufficient RAM to adequately handle all the applications that you may be working with in any one given point in time. You also have to take into consideration all of the background tasks that are working any given point in time. For example, if your antivirus is running a scan or if you are manually defragging your hard drives, this will reduce the amount of RAM as well as decreased the speed and performance of Windows itself. The end result is that DNS also slows down. In addition, and in your case, if you're dictating for long periods of time, particularly into a single application like Microsoft Word, Windows gradually requires more and more resources (RAM) to handle the amount of text that you're entering into a document. Over time, this can slow DNS down because it slows Windows down. With less than 2 GB of RAM, the result will eventually be that Windows has to swap out something to the pagefile in order to provide sufficient RAM for everything that is currently being used (active).

On the other hand, once you have at least 2 GB of RAM, and preferably 3 GB of RAM in Windows Vista because Vista is more resource intensive than XP, increasing your RAM generally does not improve the performance of Windows, and therefore the performance of DNS. However, if you change your system configuration to one that has, as in the current Intel Core2™ Duo E8400 chips or the Core2™ Extreme Quad Core QX9650 quad core processors, you are now working with 6 MB to 12 MB of L2 cache, as opposed to your Athlon which is only running 512 kilobytes of L2 cache. Don't be deceived by AMD systems that are running 2 MB of L3 cache because the L3 cache is solely for the purpose of storing interim results to be passed to applications and doesn't really have as much of an impact as L2 cache does. The current AMD processors are simply less efficient than the current Intel processors, and by current I mean most recently introduced. This includes the new AMD quad core processor. There was a previous post on another forum that disagreed with my analysis of L2 cache and DNS performance referencing a blog by an individual who is known to be well-versed in processor performance. However, none of the benchmarks used in that analysis have anything to do with the way that DNS or any speech recognition performs. The result is that their claims to the contrary relative to mine end up being a non sequitur. That is, they weren't applicable to the way that DNS works.

Lastly, once you have reached the 2 GB level of RAM, the only way to increase the speed and performance of DNS is to upgrade your system to one of the more current processor sets. Keep in mind that AMD references to speed (+3500) is not a reference to the true processor speed but to the enhancement to true processor speed based on their processor enhancement technology. You can see this if you go into your BIOS because the BIOS will tell you what the actual processor speed is. AMD is typically anywhere from 500 to 800 MHz slower in terms of true processor speed provided by the technical specifications for any particular CPU. Currently, AMD does not compete well with the latest Intel chips, particularly as regards to DNS.

Nevertheless, your problem is not a RAM problem, it's a loss of Windows performance because of the increasing amount of resources required by Windows when dictating over long periods of time. There are a couple of things you can do to reduce this propensity to slow down the transcription process under your conditions.

1. Rerun the Audio Setup Wizard as you currently are doing.

2. Periodically close your user and relaunch. This has the effect of flushing buffers and clearing out the RAM that may be becoming fragmented as a result of your continued dictation. Particularly in DNS 9/9.1/9.5, the amount of RAM and resources used by DNS gradually increases over time starting at about 180 MB an increasing to as much as 350 MB depending upon whether or not you're using DNS Preferred or one of the professional series (Legal, Medical, Professional). Since RAM is like your hard drive, the assignment of RAM to a particular task can become fragmented. However, don't use of RAM defragger program like some users attempt to do because this only exacerbates problems that you can have with DNS. It is much simpler and more effective to simply close your user and relaunch it after dictating for about two or 3 hours or when you detect a slowdown, or are you seem to be finding that your accuracy is being affected (although this is more likely due to fatigue and dictating over a long period of time which changes, or tends to change, your dictation style and quality of your enunciation/pronunciation).

3. Save your working data (documents) and close those applications, close and restart Dragon NaturallySpeaking and then reopen your applications. This is a drastic move, although less drastic than rebooting, but this tends to free up all memory so that every application has to be reassigned its own specific set of memory addresses. Because DNS must be loaded before Select-and-Say enabled applications into which you want to dictate, DNS gets for shot for memory allocation.

These techniques tend to refresh the way RAM is allocated and by applying any or all of these you will find that your dictation returns to a more stable and more accurate. Also, keep in mind that most dictation errors (when they increase) are a result of either less than optimal hardware or changes in your dictation style, such as mumbling, slurring your words, running words together, let's clear enunciation and/or pronunciation, etc.

Basically, it's important that you remember that the way you hear yourself is not necessarily the way that DNS hears you. I hate putting it that way because it tends to anthropomorphize speech recognition, which is far from the case. However, it does help to explain, in most cases, why dictation accuracy degrades over time. Consider it in this way. You do not sound to your self the way you sound to others. As a result, you can't rely on what you hear in your own dictation in order to understand what may be going on behind the scenes. The only way that you can do this is to dictate several pages saving the audio (you can only do this with the professional versions) so that you can listen to your dictation over a longer period of time by playing it back completely from beginning to end. This helps you to understand how your dictation changes over time and what possible problems your style may be creating as a result of fatigue or strain from dictating for several hours. This technique is helpful because you can review the text while you listen to your dictation, and often can point out where you are making your mistakes. It's not possible to do this effectively with short bursts of playback (i.e., a couple of sentences or a paragraph, or in the correction window/spell dialog).

Chuck Runquist
Former Dragon NaturallySpeaking SDK & Senior Technical Solutions PM for DNS

"Any sufficiently advanced technology is indistinguishable from magic." -- In memoriam of Arthur C. Clarke who passed on this week. (1917-2008)