Conducted March 12th, 2007 by RJ Sharp, InquireTec
| 1000 | A | 92 | 100 | 96 | 96.0 | 0.45 secs |
| 1000 | B | 88 | 76 | 80 | 81.3 | 0.62 secs |
| 2000 | A | 80 | 92 | 88 | 86.7 | 0.55 secs |
| 2000 | B | 84 | 76 | 76 | 78.7 | 0.73 secs |
| 3000 | A | 96 | 88 | 92 | 92.0 | 0.67 secs |
| 3000 | B | 48 | 64 | 76 | 62.7 | 0.78 secs |
| 5000 | A | 72 | 80 | 80 | 77.3 | 1.15 secs |
|
The purpose of this casual test was to learn if Apple's Speech Recognition Manager could accurately respond to grammars containing more than a few dozen items, a commonly held guideline. Testers were given 25 random names to speak out of an active language model of 1000-5000 names (LM Size). An item "passed" only if it was accurately recognized within 5 seconds. Each tester performed three tests for each language model size.
Tester A was an adult male, experienced speech user, native English speaker. This test was performed on a PowerPC G4 eMac, 1 GHz, 768 MD SDRAM, Mac OS X.4.8. The microphone was a VoiceTracker™ USB array by Acoustic Magic. The test application was built using 4D 2004.5 and Inquire4D SUI Pack v3.0 Alpha. The final test with 5000 names was to challenge certain statements made in the 1996 MacTech article, The Speech Recognition Manager Revealed, in the hopes that this barrier could be broken given the many advances made in PC hardware. The full meta data is available here. Items marked "SKIPPED" were either rejected utterances, or searches > 5 seconds. The test names were assembled from a ranked list of common first and last names obtained from the US Census Bureau website. For comments/questions about this test, email me at rj.sharp@inquiretec.com |
||||||
