A Generic, Scalable Architecture For A Large Acoustic Model And Large Vocabulary Speech Recognition Accelerator.