How to increase minimum voice recognition confidence level to mitigate voice commands triggering unintentionally?

MRTK Development
,
1

Unity Editor version: 2022.3.45
ML2 OS version: 1.10.0 (B3E.240905.12-R.038)
MLSDK version: magicleap.unitysdk 2.5.0
MRTK version: MRTK3 and magicleap.mrtk3 1.3.0

We are experiencing an issue where voice commands are triggered unexpectedly. For example, we have a voice command with keyword 'What can I say" that triggers even if we deliberately say something else like "What can you say" with emphasis on 'you'.

Consequently, I wanted to test if increasing the minimum confidence level would help.

When using the 'MRTK Windows KeywordRecognition Subsystem' we would assign a WindowsKeywordRecognitionSubsystemConfig' asset and set the 'ConfidenceLevel' property, but for 'MagicLeap Subsystem for Keyword Recognition API' there's no such option.

Any suggestions on how we can set the recognition confidence level?

2

Increasing the number of voice intents on the Magic Leap 2 will naturally reduce their sensitivity, making accidental activations less likely. However, we do not offer an API to manually adjust the activation threshold for voice intents.

3

Thanks @kbabilinski.

Do you know why they become less sensitive when increasing the number of voice intents? Do the keywords have to be similar or can they be anything?

Is there no negative performance impact when having max 100 intents vs only 10 intents?

4

Detection may slow down when using 100 intents, but we haven’t received any negative feedback from users running this many. The system compares spoken phrases against the registered intents and selects the closest match. When more intents are registered, the system has to be more selective, which can reduce accidental activations but may slightly impact the speed at which a phrase is triggered.

The keywords don’t necessarily have to be similar for this effect to occur. The voice intent system still evaluates all registered phrases, but with a larger list, it must be more discerning in choosing the best match.

I’ll submit your posts as a feature request to our Voice of Customer program to explore potential improvements in this area. However, this feature is not currently on our roadmap.

1 Like