Has Voice Command Detection Improved in MRTK3?

[YonatanMARS]

Hi everyone,

I'm currently evaluating the transition from MRTK2 to MRTK3, and I wanted to ask specifically about the voice command (keyword) recognition system.

In MRTK2, we used KeywordRecognizer via the Speech Commands Profile, which worked but often struggled with:

• Accent sensitivity, especially for non-native English speakers
• Latency between speaking and detection
• Occasional really large latency (almost 1 minute) when the app is under heavy load

Now that MRTK3 introduces the new WindowsKeywordRecognitionSubsystem, which can optionally use SelectKeywordRecognizer via OpenXR 1.5+, I’m wondering:

• Has the recognition accuracy improved, particularly for users with non-native English accents?
• Is the response time (latency) better or worse than the old Unity-based KeywordRecognizer?
• Are there any configuration tips (e.g. model packs, confidence levels) to improve detection quality?

I'd love to hear feedback from anyone who's tested both MRTK2 and MRTK3 voice features, especially on HoloLens 2 with OpenXR.

Thanks in advance!