Improving recognition accuracy in NaturallySpeaking
Microphones, sound cards, etc. can cause errors. This is particularly true on older laptop systems where components are crowded. Listen to your system! Listen to the playback to determine if a clean sound is getting to the system.
With laptops, sometimes accuracy is much better running off battery than off of the A/C. That is due to interference from the AC-DC converter, a common problem. Using a USB adapter with your microphone will often resolve this problem.
Be consistent with your microphone. Do a recording, moving your microphone while you speak. With most microphones, you'll notice differences in sound levels and clarity of the sound.
The following are suggestions for improving recognition accuracy in Dragon NaturallySpeaking 5. As accuracy improves, the response speed of the speech recognition software should also improve.
Consider the following suggestions to improve recognition accuracy.
- Sound Card: A good sound card is the most important part of the computer system when running Dragon NaturallySpeaking. Unfortunately, it is sometimes difficult to define what makes a good sound card.
For NaturallySpeaking and speech recognition in general, the important part of the sound card is the input channel. This channel includes the microphone input, amplifiers, and A/D converters. The fact that a sound card produces good output does not mean that the sound card produces good input. Checking the sound card specifications will help to determine the quality of the sound card.
- Microphone Placement: To improve recognition accuracy, review the placement of the microphone. Incorrect placement typically causes additional small words like "to" and "a" to be added to the sentences. This usually means the microphone is in the path of the air stream and should be a little higher or positioned off the corner of your mouth. Make sure that the microphone is also consistently located during the dictation session and in subsequent sessions. Consistent microphone placement in the correct position is essential for high accuracy.
- Pauses Between Phrases: Another possible cause for the insertion of extra small words into dictation, or the incorrect splitting of larger words into smaller words, is an incorrect setting for the "Pause between phrases" option. If the set time is too short, recognition can be affected. If you experience the above-mentioned problem, lengthen the pause between phrases setting. Change this setting in the following manner:
- From the Tools menu, choose Options.
- Select the "Miscellaneous" tab.
- Increase the Pause Between Phrases setting. (The default pause setting is 250 milliseconds.) If the setting is less than this default setting, it may have been decreased for faster response.
- Dictating Tips: Consider these suggestions when using Dragon NaturallySpeaking.
- Verbalize every word, avoiding slurred pronunciation or swallowing of sounds. (For example: Do not pronounce the word "to" like " tuh.") For NaturallySpeaking to recognize a word, say the word. In conversations between people, speakers can sometimes drop words or slur words together. If a word is dropped, do not expect NaturallySpeaking to find it. That said, NaturallySpeaking does include shortened pronunciations of some small function words like "n" for "and."
- Speak at a normal pace. Do not talk slowly to NaturallySpeaking even if the computer is slow. Talking slower will actually degrade recognition accuracy. Practice by reading out loud. This will come closest to the best way to dictate to Dragon NaturallySpeaking.
- Speak continuously. NaturallySpeaking recognizes longer utterances much better than shorter ones. Avoid speaking one word at a time since this does not improve accuracy.
- Speak the way you trained. When you ran General Training, you spoke a certain way and Dragon NaturallySpeaking learned the way you talk from those samples. If you speak like you are reading something, your accuracy may improve. Similarly, you can rerun General Training, further adapting NaturallySpeaking to the way you dictate.
- Think first, talk second. Composing while dictating is a new skill for many people. The trick is to think ahead. Try to think about what you are going to say before you say it. Then, when you say it, your dictation will be clearer because you already know what you are going to say. If you talk without first thinking about it, you will find that your words get mumbled as you change your mind while you speak. Dragon NaturallySpeaking will understand "natural" speech but not "conversational" speech. Pretend that you are a broadcast journalist reading the news from behind an anchor desk as opposed to talking to your friend over coffee.
- Correct Misrecognized Words: For best results, select the word just before and just after the misrecognized word as well, and correct them as a phrase from the Correction dialog box. Also, train the word or words that were recognized instead of what you intended.
Note: If you do not use the Correction dialog box, recognition accuracy will not improve.
- Run the Vocabulary Builder: (Formerly called "Topic Builder" in NaturallySpeaking Deluxe v.2.02.) Vocabulary Builder will add words you tend to use to the active vocabulary and update its word usage modeling.
- Delete Unused Word: If a word you never use is recognized for a word you use often, delete the unused word in the Vocabulary Editor.
- Acronym Pronunciation: If you have trouble with acronyms, use the Vocabulary Editor to assign it an alternative pronunciation. In the Spoken Form field, enter the words you will say; in the Written Form field, type the word as it will print. For example: If you say "USA" and it comes up as "usually", you could write the spoken form as "American acronym" and the written form as "USA."
This technique also works for "sounds-like" pronunciations. For example: If you pronounce the word "record" as " rickert", use Vocabulary Editor to enter the spoken form of the word "record" as " rickert."
- Run Additional General Training: From the Tools menu, run an additional general training procedure, especially if the background noise in your environment has changed, or if you have a cold.
- The trouble with Commands: To allow the spoken commands to execute as desired, press and hold the CTRL key and issue the command or commands. This will prevent the program from typing the spoken command.
Additionally, when a command is recognized as something else, you can use the Correction Dialog to correct the phrase to its constituent words. When you do this, the words will be entered into your document on confirmation, but they can then be deleted and the command should be more accurately recognized. Consider the following example:
- You said, "Go To Sleep."
- NaturallySpeaking typed, "gouda sleep."
- Select the words "gouda sleep" and, using the Correction Dialog, correct it to read "Go To Sleep."
- Click OK. The command will be better recognized next time. (You can delete the typed words "Go To Sleep" at that point.)
Note: Remember that commands require a brief pause before and after they are spoken to distinguish them from dictation words. If the length of this pause appears to be a cause of recognition difficulty, try adjusting the "Pause Between Phrases" slider on the Miscellaneous tab of the Options dialog. See Step 3 above.
- Save Speech Files: After correcting misrecognitions and training words, be sure to save your speech files before closing NaturallySpeaking.
If you spell in the Correction Dialog box, use the alpha-bravo "military" alphabet.
Improving Dragon NaturallySpeaking accuracy on your M-Tech Computer
The Accuracy Center provides "one-stop shopping" for all the tools you can use to improve your accuracy and add words to your vocabulary. Say "display Accuracy Center" or click Accuracy Center on the DragonBar Tools menu.
If you are not sure which of the Dragon NaturallySpeaking tools you need to use, let the Accuracy Center help you. With the Accuracy Center open, click or say "which tool should I use?" to start the Accuracy Assistant.
The best way to improve recognition accuracy is to speak clearly and enunciate each word. You do not have to speak loudly or slowly, just avoid mumbling or slurring your words.
You can optimize recognition accuracy by speaking in long, well-enunciated phrases or sentences. Speaking slowly and deliberately, in short phrases or single words, can actually result in more recognition errors. Longer phrases provide more context, which helps Dragon NaturallySpeaking recognize individual words.
If a particular word is regularly mis-recognized or you want to give the word an alternate pronunciation, use the Training dialog box to train it. To open the Training dialog box say "train words" or click Train on the DragonBar Words menu.
If you change audio hardware or move to another location, you should run the Audio Setup Wizard again. You do not have to create a new user. To open the Audio Setup Wizard, open the Accuracy Center and either click or say "check your audio settings."
Use the Increase Accuracy from Email tool to add the names of your email contacts and to teach Dragon NaturallySpeaking about your personal writing style and vocabulary by analyzing your sent email messages. This can significantly improve recognition accuracy. This tool is in the Accuracy Center. Say "display Accuracy Center" and then say "increase accuracy from email."
A good way to find out whether or not a word is in the active vocabulary is to type it in the Vocabulary Editor. (Say "Open-Vocabulary Editor" or click View/Edit on the DragonBar Words menu.) If the word appears in the list, it is already in vocabulary; if it does not appear, you can click the Add button to add it.
You can view only the words you added to the vocabulary by saying " Open-Vocabulary Editor" or clicking View/Edit it on the DragonBar Words menu and then selecting Show Custom Words Only from the Display list in the Vocabulary Editor.
When you correct a misrecognized word, the program is less likely to misrecognize that word in the future.
You can improve recognition accuracy by adding multiple-word phrases with unusual capitalization (e.g., "The Mayberry Tribune") to your vocabulary. Say "add new word" or click New on the DragonBar Words menu.
You can see a list of the last few recognition results. Say "view Recognition History " or click View Recognition History on the DragonBar Words menu.
Run a spelling check on your documents before processing them in the "Add Words from Your Documents" dialog box. This prevents misspelled words from being added to your vocabulary.
You can create a Spoken Form for words you prefer not to say aloud. For example, you can create a text command that types "781-555-1212 " whenever you say "my phone number."
It is normal for words in the Results box to change sometimes while you are dictating. Dragon NaturallySpeaking may better recognize what you said after you complete a phrase or sentence.
How to improve accuracy with Dragon NaturallySpeaking Recorder Edition
There are multiple ways to improve accuracy when using Dragon NaturallySpeaking Recorder Edition:
- Running the Acoustic and Language Model Optimizer.
- Performing corrections.
- Analyzing e-mail.
Method 1: Performing corrections
After the user transcribes an audio file, the user should correct any words that were not recognized properly. With the Recorder Edition, it is not possible to correct words by voice.
To perform corrections:
- Open a digital recorder user profile.
- Perform a transcription of an audio file by using the "Transcribe" button on the DragonBar.
- Using the mouse, highlight the text that needs to be corrected.
- Click the "Correction" button on the DragonBar or press the Correction Hotkey (the "minus" key on the numeric keypad, by default) to display the Correction dialog.
- If the correct word is displayed in the Correction dialog, select that word to make the correction.
Note: The user can click the "Playback" button at this time to verify what was originally dictated.
- If the correct word is not displayed in the Correction dialog, click the "Spell That…” option and type out the correct word in the Spell dialog.
- Press the "Enter" key or click "OK” to make the correction.
- Correcting in this manner will adapt the user's voice files, leading to continuously improving accuracy.
Method 2: Analyzing e-mail
Use the Increase Accuracy from E-mail tool to add the names of people the user has sent an e-mail to into the Dragon vocabulary. The user has the option to edit these names and can also have Dragon scan the contents of the user's sent e-mail and add it to the language model.
To analyze e-mail:
- Open a digital recorder user profile.
- Click "Tools > Accuracy Center".
- Click the "Increase accuracy from e-mail" link.
- The Increase Accuracy from E-mail wizard opens. Click the "Next" button.
- Select the e-mail source to be used:
- Lotus Notes
- Microsoft Outlook
- Outlook Express
- Select one or both of the following options:
- Add contact names from e-mail to the vocabulary
- Improve my speech files from my e-mail writing style
- Select one of the following options:
- Consider e-mail written since my last scan
- Consider all e-mail
- Click the "Next" button to start analyzing e-mail.
Method 3: The Acoustic and Language Model Optimizer
Before running the Acoustic and Language Model Optimizer, make sure the "Store corrections in archive" option is enabled under "Tools > Options > Data tab". If this option is not enabled, the user may see a message that there is not enough data for the Acoustic and Language Model Optimizer to process.
To run the Acoustic and Language Model Optimizer:
- Open a digital recorder user profile.
- Click "Tools > Accuracy Center".
- Click the "Run the Acoustic and Language Model Optimizer" link.
- Enable the "Perform Acoustic Optimization" and/or the "Perform Language Model Optimization" options.
- Select "Perform Acoustic Optimization" to update the user profile with accumulated acoustic data from any corrections made by the user.
- Select "Perform Language Model Optimization" to update the language model. The language model contains statistical information that predicts which words are most likely to occur in the context of the user’s speech.
- Click the "Go" button to begin the optimization.
- The Acoustic and Language Model Optimizer dialog box will display the amount of time to finish adapting the user files and display a progress bar to indicate how much of the job is complete. It generally takes as long to adapt the user files as the amount of data stored. For example, if there are 30 minutes of acoustic data stored, the processing time will be approximately 30 minutes.
- Once the optimization has completed, save the user profile by clicking "NaturallySpeaking > Save User Files".
General tips for increasing accuracy:
- If using the digital recording device's built-in microphone, always hold the device at the same distance from the user's mouth. Some users rest the device on their jaw for the most consistent results.
- Consider connecting a headset microphone to the digital recording device rather than using the built-in microphone. Using a high-quality external microphone is likely to produce better quality recordings.
- Adjust the Speed vs. Accuracy setting under "Tools > Options > Miscellaneous tab". The Speed vs. Accuracy setting adjusts the number of words and phrases that the program examines in order to find a word match. The more words that are examined, the greater the possibility of an accurate match. However, this comes at the expense of using more memory and taking a long time to recognize the user's speech. The user can adjust the Speed vs. Accuracy slider to influence the performance of Dragon. As a general rule, the user should move the slider farther to the right on faster machines than on slower ones.
Recognition starts to deteriorate after dictating for awhile.
Several factors could cause this problem:
- The microphone settings might have been changed. Please run the Audio Setup Assistant. This will readjust your microphone settings so that ViaVoice will work correctly.
- You might not be using your own voice model. Make sure you use the correct username that contains your own personal language model. You can use the User settings to change users.
- You might be positioning your microphone incorrectly. The microphone boom should be angled toward the corner of your mouth (not up or down) so that the flat part on the microphone element under the foam faces the corner of your mouth. Audio Setup Assistant illustrates the correct positioning of the microphone.
- Make sure that you are not breathing directly into the microphone mouthpiece or speaking too loudly when dictating. Also, make sure the foam microphone wind screen is attached and in place.
Tips for better performance from Dragon NaturallySpeaking on your M-Tech Computer
- For best results, creating a User profile should be done in your "normal " environment, with the same background noise as when you will be dictating. Recognition accuracy can sometimes be affected by the presence of air conditioners, fax machines, printers"¦
- Version 10 of Dragon contains special acoustic models for a number of "accents " heard in the US (this is particularly important for natives of the UK or of the Indian sub-continent). When creating a user profile, you will be able to choose an "Accent "; if unsure, just pick the "General " accent.
When creating a User profile, Dragon presents you with a list of dictation sources. If you are not using the default (microphone plugged into the Mic-In jack), be sure to indicate your device's type: for instance, USB Microphone or Bluetooth Microphone. (The Help provides details on using special dictation sources, including digital recorders.)
If you already have a user profile and want to use an additional dictation source, add a source instead of creating a separate profile for the new device (use the Open User dialog's Source button.) That way, the same Vocabulary will be used
"" and refined
"" whether you are using your USB microphone, Bluetooth microphone, digital recorder, etc.; remember, ongoing vocabulary personalization is one of the keys to getting productivity and convenience in speech recognition!
Keep the microphone about a thumb's width away from the corner of your mouth, not directly in front of it. It should not touch your skin, lips, or hair. Experiment with the best distance for your particular situation, and gently bend the headset's "boom " if necessary in order to have the microphone in the right place.
Noise-canceling microphones only listen on one side. Check that the listening side (often marked by a dot or the word TALK) is parallel to your face and not tilted up or down.
Position the microphone carefully and consistently every time you use it. If you notice the accuracy decreasing, it might be because the microphone has moved out of position.
If you want to test your microphone independently of Dragon, you can use the operating system's Sound Recorder "" accessible from the Windows Start menu > Programs > Accessories.
If your microphone model has a mute switch, make sure it is in the ON position before you start to speak.
When you read a text for Dragon's acoustic training (whether when you first "enroll ", or later, if you wish to boost accuracy), Dragon learns how you sound when you dictate; using the same tone, pace, and volume you are likely to use will help it recognize your future dictations most accurately. Saying punctuation isn't necessary here, but it's a good idea to say at least the periods, commas, and colons, to get into the habit.
Here is a tip from one of our customers who happens to be as much an expert on Dragon usage as you can get.
. I find that on my "High Horsepower" machine, that setting the"Accuracy Slider" in "Tools/Options/Miscellaneous" to about 25%-40% as shown below, I get about the same accuracy, but the response time of the program is much faster. I could probably go even lower if I wanted to be aggressive. I have just not tried to see how low I could push it. If I set the slider to "Most Accurate", DNS doesn't perform any better, for practical purposes, and the response time goes way up. Processing time doesn't seem to be affected as much as latency, e.g., if you dictate a paragraph, it comes back in about the same amount of time. If you dictate two words, it takes forever to come back compared to lowering the accuracy slider.
Here is a set of links for Rev. B for Dragon 13. As it turns out, DNS 13 is giving many users troubles in a variety of areas. We will attempt to keep these links up to date. Lunis at KnowBrainer.com has been kind enough to keep M-Tech and its customers in the loop on this. Many thanks to Lunis and the gang over at KnowBrainer.com
Nuance has been working on this for a couple of months. I have now uploaded REV B for everything. Here are the links if you have a need
DNS Home Edition 13
DNS Premium 13
DNS Professional 13
DNS Legal 13
Here is download copy of KnowBrainers Recommended Tweaks for Dragon. Review them and the pros and cons of each tweak before making the adjustment.