Apologies for the delay in answering.
There are quite a few components working together here.
First and foremost: Null Rom. Without that, nothing else happens.
Second: offline voice recognition
This was pretty much taking all the language files from my phone and transplanting them to the watch, minding permissions.
Third: AutoVoice and Tasker
These are market apps and are fantastic for automation. You'll need them on both phone and watch. I'm using two profiles for this. The first turns on AutoVoice recognition in continuous mode whenever the screen turns on, and off when the screen goes off. The second profile is an AutoVoice recognize with "galaxy" as the command filter. It initiates an intent with the rest of what I say as a data payload.
Fourth: M2D manager
This is also available on the market, and needs to be on both devices. It is a Bluetooth bridge for Android intents. Tasker sends out an intent formed for M2D with the voice command as data. M2D transmitted it to the phone where it seems out a specified intent, still containing the command. Tasker listens for that intent. Once it had it, I use the AutoVoice test feature to send the command text in as if it had been spoken to the phone, so I can use all of my previously written voice control profiles. M2D also works the other way around, which I take advantage of for notifications and such.
Sent from my XT1060 using XDA Premium 4 mobile app