Using speech diagnostics
Use the speech diagnostics tool to determine what the Speakeasy engine recognizes successfully, and manually adjust the detection threshold and confidence level to improve recognition.
Before you begin using speech diagnostics, make sure the headset is connected and working properly.
Speech diagnostics is designed for manually changing settings or for diagnosing problems.
Absolute threshold
The absolute threshold is how much energy is required before Speakeasy recognizes the input as speech. If the Speakeasy engine isn't returning results, the microphone input may not be loud enough for the engine to recognize speech. Lower the absolute threshold to detect more noises as speech. If the Speakeasy engine is returning speech results when there is background noise, raise the threshold to detect fewer noises as speech.
Confidence values
When the speech-to-text engine detects speech, it tries to match it to the expected phrases in the available grammar files. The degree to which the spoken phrase and an expected response match is called the confidence value. A higher confidence value indicates a better match. If the confidence value is too low, the input is rejected and Speakeasy doesn't return a result.
In the speech diagnostics tool, when the confidence value for detected speech is above the threshold, the term is displayed in green at the bottom of the screen. When the confidence value for a term is below the confidence threshold, Speakeasy displays the best match in red at the bottom of the screen. To view other terms from the grammar file that may have matched the detected speech, click on the term to see the other phrases that Speakeasy tried to match and the confidence values associated with them.
The speech diagnostics tool shows the confidence value for each phrase that Speakeasy detects, and allows you to configure the confidence threshold or value. If the threshold is too low, Speakeasy may recognize too many false positives. However, if the threshold is too high, Speakeasy may not recognize accurate phrases.
Troubleshooting speech
To use the speech diagnostics for troubleshooting poor recognition, you will need to be able to reproduce the speech – and the background noise as well, if that is a factor. You may want to create a recording of the user speaking the phrase several times, and then play the recording back while the speech diagnostics are open.
Start by adjusting the absolute threshold or confidence level to see if you get better results. If changing these settings improves results, you can use a script to set them automatically. For more information about using scripting to change Speakeasy settings, see the Voice APIs in the Velocity Scripting Guide.
If there is a term that Speakeasy consistently fails to recognize, or it consistently returns the wrong result, consider changing the term in the grammar file to make it more distinct. To make terms more distinct, you can:
•Use a different phrase. For example, instead of 'faster,' use 'warp speed.'
•Add another word to the term. For example, instead of 'faster,' use 'speak faster.' Or instead of 'two', use 'two ready.'
•Spell the word more phonetically in the grammar file. For example, instead of 'picking,' use 'peeking.'
For information on specifying terms in grammar files, see the Speakeasy Reference Guide.
To use speech diagnostics:
•To open the speech diagnostics in the Velocity Client, navigate to Menu > Speech Diagnostics.
•To test against a specific set of phrases, use the Grammar dropdown menu to select the grammar file.
•To pause speech recognition, set the Speech Detection option at the top of the page to Off.
•To adjust the detection so that Speakeasy detects speech with a lower energy level, use the Absolute Threshold slider. The green bars at the top of the page show the current detected energy level and the average energy level for the past 5 seconds. Try to set the absolute threshold to the highest threshold at which you consistently get correct results. Setting the energy threshold too low means that Speakeasy detects more background noise, and may return more false positives.
•To adjust the confidence threshold, use the Confidence Level slider. Try to set the confidence level to the highest threshold at which you consistently get correct results.