On device speech models optimization and deployment