Optimizing SphinxCMU Project Home Page

Optimizing SphinxCMU – Multithreading, SIMD and Microarchitectural Considerations

Adeen Flinker (sadeen@t2.technion.ac.il)

Ruth Bernstein (sruthb@t2.technion.ac.il)

Supervisor – Zvika Guz (zguz@tx.technion.ac.il)

Technion - Department of Electrical Engineering - Software Lab

May 2005

Over the course of this project, the SphinxCMU speech recognition engine was optimized using several technologies.

	Multithreading support was implemented using a data composition strategy (both OpenMP and Windows API were used and assessed separately).
	SSE2 instructions (SIMD) were introduced in critical code sections.
	Various diagnostic tools were used in order to gather optimization information and subsequent code changes were incorporated

The final code showed a 30% speed up over the original code.