Merge "Don't unload sound model in start recognition unless the model changes" into lmp-dev