SPEC CPU2006 Benchmark Description

Benchmark Name


Benchmark Author

Sphinx Speech Group, Carnegie Mellon University. Thank you especially to CMU Researchers Evandro Gouevea, Arthur Chan, and Richard Stern for their assistance to SPEC in creating this version. Thanks also to Paul Lamere of Sun for timely consulting on many porting questions.

Benchmark Program General Category

Speech Recognition

Benchmark Description

Sphinx-3 is a widely known speech recognition system from Carnegie Mellon University.

This description assumes that the reader has already seen the Sphinx 3 introduction, which provides an excellent introduction to the inputs, outputs, and operation of the code. (A copy of this file as of mid-2005 is included in the SPEC CPU2006 kit, as $SPEC/benchspec/CPU2006/482.sphinx3/Docs/sphinx3-intro-CMU.html.)

CMU supplies a program known as livepretend, which decodes utterances in batch mode, but otherwise operates as if it were decoding a live human. In particular, it starts from raw audio, not from an intermediate (cepstra) format.

Although in real life IO efficiency is obviously important to any speech recognition system, for SPEC CPU purposes we wish to concentrate on the CPU-intensive portions of the task. Therefore, main_live_pretend.c has been adapted as spec_main_live_pretend.c, which reads all the inputs during initialization and then processes them repeatedly with different settings for the "beams" (the probabilities that are used to prune the set of active hypotheses at each recognition step).

Input Description

The AN4 Database from CMU is used. The raw audio format files are used in either big endian or little endian form (depending on the current machine).

Output Description

Correct recognition is determined by examination of which utterances were recognized (see lines "FWDVIT" in the generated .log files), as well as a trace of language and acoustic scores.

Programming Language


Known portability issues



Last updated: $Date: 2011-08-16 18:23:17 -0400 (Tue, 16 Aug 2011) $