For use of Google's voice recognition API, see Omni's Launcher3 commit history.
For generic offline recognition - not sure where to look.
Qualcomm's voicewakeup is designed to do ONE thing and do it well - wake up the device in response to a voice command. All other voice-driven features are done using different software/approaches than Snapdragon Voice Activation.
It's been kind of mentioned previously in a few other places (including, I think, earlier comments here...) that SVA is tightly integrated with a number of specific Qualcomm features:
1) The audio chip can be put into a mode where it issues an interrupt to the host CPU when the input level exceeds a certain value. (e.g. only process loud sounds)
2) The hexagon DSP is very power-efficient and can process data from audio chip wakeups efficiently. With SVA - interrupts from the audio codec don't wake the apps CPUs, they wake up the Hexagon instead.
3) Only once the Hexagon matches a predefined wakeup pattern does it wake up the host to do any further processing.
e.g. SVA does not solve the "what did the user say?" problem, it solves the very specific "Did the user say X? If they said anything other than X I don't care" problem.