Sound Input
One of the best microphones for speech recognition uses active noise canceling technology. Purchasers often contact us and say, "Microphone sound on the audio setup test is in the high part of acceptable, sound quality is at the bottom of acceptable. What is wrong with my microphone?" Our first question is, "Have you tried dictating with the microphone?" The answer is usually no. The user is invariably pleased with the improved accuracy upon dictating. The problem was not with the microphone and not with the test. The audio setup was developed to test for a specific type of microphone. Microphones supplied with inexpensive speech recognition software packages are in and of themselves inexpensive. In a quiet environment the microphones work well enough to demonstrate the potential of the speech recognition software. The test in the speech recognition software is only as good as the equipment it was designed to test. After testing dozens of microphones a year we developed a testing method to take into account the variables that make a microphone a high performer. Items of importance are:
- Sound input quality
- Accuracy of the delivered sound
- Background noise elimination
- Ergonomic ease of use and comfort
SOUNDCARD/MICROPHONE QUALITY
The Windows Sound Recorder is a useful applet to test a soundcard and microphone for unwanted electronic noise. Open Sound Recorder, click File, click Convert Now, drop down the list with the heading Name and select CD- quality. Hit OK. CD-quality is only important for purposes of this test. Speech recognition software automatically uses the audio format it was designed for. Make a recording. Play back the recording through good speakers or a good earphone. If there is no electronic noise or static you can assume your soundcard and microphone are capable of delivering clear sound without extraneous noise.
ACCURACY TEST By reading a standard 100-utterance paragraph trained to the speech recognition software, accuracy is easily measured in terms of a percentage. Utterances include a text macro, commands such as new paragraph, period and comma. Don't ask us for our paragraph: Make up your own. The test begins by creating and processing an enrollment training. The standard paragraph is then read in a very quiet environment. Any good microphone should have five or fewer mistakes the first reading. Correct misrecognized words and read the paragraph again. A good microphone should have no more than two misrecognitions at this point.
UNWANTED NOISE REJECTION Progressively add sound (recorded office noise or music) from speakers to the front and rear. If a tested microphone routinely keeps accuracy in the 98 percent99 percent on the standard paragraph, the test continues till we reach the noise level of a typical open cubicle office. If the microphone performs well at this point, it is deemed accurate and noise canceling.
ERGONOMICS Is the microphone comfortable after being worn or held for a typical dictation session? Is it is easy to put on if a headset, or easy to hold for a handheld? What features does it have? Sometimes a microphone may have no features. This is a feature within itself. We give several clients the microphone to use for a week. Feedback on how they work in real life office conditions is a major consideration. When a microphone gets an "A" grade from these tests, we are confident it will meet most people's needs.
Martin Markoe has been involved in the speech technology field since 1994. He may be reached by e-mail at martin@speechcontrol.com