Fixed-Point Arithmetic

There are two main requirements for embedded/mobile systems: one is low power consumption for long battery life and miniaturization, the other is low unit cost for components produced in very large numbers (cell phones, set-top boxes). Both requirements are addressed by CPU’s with integer-only arithmetic units which motivate the fixed-point arithmetic implementation of automatic speech recognition (ASR) algorithms. Large vocabulary continuous speech recognition (LVCSR) can greatly enhance the usability of devices, whose small size and typical on-the-go use hinder more traditional interfaces. The increasing computational power of embedded CPU’s will soon allow real-time LVCSR on portable and lowcost devices. This chapter reviews problems concerning the fixed-point implementation of ASR algorithms and it presents fixed-point methods yielding the same recognition accuracy of the floating-point algorithms. In particular, the chapter illustrates a practical approach to the implementation of the frame-synchronous beam-search Viterbi decoder, N-grams language models, HMM likelihood computation and mel-cepstrum front-end. The fixed-point recognizer is shown to be as accurate as the floating-point recognizer in several LVCSR experiments, on the DARPA Switchboard task, and on an AT&T proprietary task, using different types of acoustic front-ends, HMM’s and language models. Experiments on the DARPA Resource Management task, using the StrongARM-1100 206 MHz and the XScale PXA270 624 MHz CPU’s show that the fixed-point implementation enables real-time performance: the floating point recognizer, with floating-point software emulation is several times slower for the same accuracy.