Performance Analysis of Speech Recognition Software

This paper characterizes the behavior of a speaker-independent large vocabulary continuous speech recognition (LVCSR) system. This system is used to dictate Chinese (Mandarin) utterances of different speakers and achieves a word recognition accuracies of 85%~96% depending on the cleanness of input signals and the complexity of the spoken sentences. Several methods are used to characterize its processing behavior. First, the same system is run on different Intel platforms and their performance measured. Second, we use an Intel performance monitoring toolset – Vtune to read hardware counters build in the CPU. These counters measured the instruction distribution as well as processor utilization rate. Third, a full platform simulator SoftSDV together with a cache simulator is used to study its memory behavior in more detail. We find this software system to have a large memory working set. Data access to first level cache has good locality. There are two groups of memory usage displacing each other in the second level cache. Second level cache miss rate declines much fast after the size increases beyond 8MB. We also identify a few instructions that cause a larger number of level-2 cache misses. Using software prefetching we improve the overall performance by an average of 7%.