The social media giant's first real-time audio model targets speech recognition as it seeks to close the gap with rivals.