Algorithm Optimizations: Low Computational Complexity

Abstract

Abstract. Advances in ASR are driven by both scientific achievements in the field and the availability of more powerful hardware. While very powerful CPUs allow us to use ever more complex algorithms in server-based large vocabulary ASR systems (e.g. in telephony applications), the capability of embedded platforms will always lag behind. Nevertheless as the popularity of ASR application grows, we can expect an increasing demand for functionality on based applications by natural language understanding (NLU) systems leads to increased vocabulary sizes and thus the need for greater CPU performance. In this chapter we present an overview of ASR decoder design options with an emphasis on techniques which are suitable for embedded platforms. One needs to keep in mind that there is no one-size-fits-all solution; specific algorithmic improvements may only be best applied to highly restricted applications or scenarios. The optimal solution can usually be achieved by making choices with respect to algorithms aimed at maximizing specific benefits for a particular platform and task. IntroductionWhile systems for dealing with large vocabulary recognition in real-time ASR have been available for many years, deployment on platforms with limited resources presents new challenges with respect to many aspects of the overall ASR system design. There is an obvious tendency to utilize any new beneficial technique over the wide spectrum of ASR applications and platforms. But in numerous respects embedded platforms lag behind the state of the art in speech recognition with respect to general CPU architectures. The limitations of the embedded platforms affecting this state of affairs includes: lower CPU clock speeds, limited or missing ability to process floating point operations, and restricted memory capacity both with respect to speed and size having implications to all levels of caching. Those constraints are bad enough, but added to this is the relatively restricted capability of tools/platforms tailored specifically to embedded systems development, which results in many important language features (e.g., templates in the C++ case) not being fully supported.Hence any new algorithm targeting embedded platforms must therefore be evaluated in light of the above restrictions. Possible actions that one needs to consider taking may include finding more efficient implementation of a given algorithm, finding and applying an even more efficient algorithm with the same or similar results, embedded platforms as well. For example, replacing simple command and control grammar-