Rss Categories

Exporting Tuner Data

Reference Number: AA-02113 Views: 6872 0 Rating/ Voters

The Speech Tuner contains a wealth of information, from raw collected data from ASR and TTS interactions that were imported, to calculated metrics and predicted estimates of optimal values.

Many of these things can be viewed in a number of useful ways when using the Speech Tuner - in particular, when using the Tuning Wizard, or the Tester View.

Often, people don't realize that much of this data can be exported in one way or another, so that for example, you could compile your own accuracy or performance reports, or perhaps you need a list of problematic grammars, or phrases. Most of these things have been designed to be exported to disk in common data formats.

On most Views within the Speech Tuner, you have a context-menu available whenever you right-click on a table or a graph. Many of these context, or popup, menus have options to "Save..." or "Export..." the list, audio or the image you clicked on.

By way of example, you can export the entire interaction list by simply right-clicking on the list as shown here:

The popup menu shown has a "Save List to File..." option. If you select this, you will be prompted to select a folder and filename where the interactions data will be written.

Note that starting with LumenVox version 14.2, a new data column was added to the Interaction List - this column is labeled "Class", short for Accuracy Classification.

To activate the Class column, you need to click on one of the Confidence Histograms to select a desired Confidence Threshold, as shown here:

In this example, the selected Confidence Threshold is 657, according to where the mouse is clicked. You can select any value you like for this Confidence Threshold Selection, which will then be used to update the Interactions List "Class" column, which will then show the accuracy classification, based on this selected Confidence Threshold, as shown here:

Note that an indication of In-Grammar (IG) as well as Out Of Grammar (OOG) is shown, since this is important when evaluating accuracy. Other classifications after the hyphen describe the specific classification, such as Correct Accept (CA), Correct Reject (CR), False Accept (FA) and False Reject (FR).

These classifications may be very useful when evaluating overall application performance when using different Confidence Threshold values. Please see our article on Calculating Speech Recognition Accuracy for more information about these terms and how they are used.

Interaction lists exported from LumenVox version 14.2 on will have a comment at the end of the data indicating the selected Confidence Threshold when the report was saved. The column information will also be exported along with the rest of the list data.