Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
You can help your end users achieve low recognition error rates by following these integration guidelines. Be aware that certain bad integration patterns can make it impossible for the system to adjust to the user's way of speaking and therefore cause increased recognition error rates.
Adaptation
Speech recognition user profiles consist of various models that represent a user's way of speaking.
When a user is using the system for the first time, initial models are created for that user. Over time, the system adapts these models to the specifics of the user's way of speaking. Via this process, the profile is personalized for the user, reducing recognition errors.
An essential element of this adaptation process is learning based on the user correcting errors. If a word is incorrectly recognized and subsequently corrected by the user, the system can learn from the mistake and correction and avoid future errors. As this is a statistical process, not every correction will immediately result in the error never occurring again. However, in general, error rates are reduced over time if recognition errors are corrected and these corrections are made available to the system.
How the system learns from corrections
To make corrections available to the system, Dragon Copilot regularly sends text from speech-enabled edit controls to the server, where it's fed into the backend adaptation processes.
The speech recognition profile of a Dragon Copilot user is continuously adapted to the user's way of speaking. The system learns by comparing the text created by Dragon Copilot to the final text after it has been corrected by the user. This is a statistical and heuristic process that significantly reduces the number of recognition errors over time.
As the user might edit the same word several times, only the 'final' text state is considered for adaptation.
For adaptation to work efficiently, it's essential that the system is provided with corrected text. To ensure this, follow the guidelines for API integration and communicate the end user guidelines to your users.
Application logic - what you must do
Make sure a user's text corrections are visible to Dragon Copilot.
Avoid programmatic text changes to dictated text, because changes to dictated text might be interpreted as manual corrections by the user.
Don't do text post-processing in edit controls while they're still speech-enabled. Don't replace dictated words by other text (e.g. to realize a feature such as 'normals expansion'); use Dragon Copilot texts or application commands instead.
Guidelines for end users
- Always correct misrecognized words. The system learns from your corrections, which means that corrected errors will appear less likely in the future. The system will assume that any misrecognized words you left in were recognized correctly, and these errors will continue to appear.