FPGA Implementation of a Pipelined Gaussian Calculation for HMM-Based Large Vocabulary Speech Recognition

Table 1

Timing results for software implementation.

	Frames/s	Speed (× real time)	Percentage of Clock cycles spent on gaussian calc.
Sequential C++	114	0.88	80%
C++ with SSE intrinsics	902	0.11	48%
8-thread parallel C++ with SSE	2756	0.036	48%