Microphones and signal quality for optimal speech recognition

Microphones make a big difference in speech recognition accuracy. Basically you need to have a 30dB signal-to-noise ratio for optimum accuracy, and many PC microphones don't do much better than 20dB in optimal operation, and sometimes much worse.

How to check for good signal quality

A good way to check the signal quality is to record a bit of audio using your setup, and look at the waveform in some waveform editor like Entropic's ESPS/Waves system; the signal should be quite flat during the silences and the speech-signal peaks should have sample values of at least +-5000 or more, even better if they are +- 10000 or 15000 in a loud vowel in the word "now". At the same time there should be no clipping, where the sample values are maxed out at +-32768; that introduces very bad distortion.

How to fix poor signal quality

If you have clipping, turn the volumes down. If peaks are too low, turn the volume up. If the noise floor is visible instead of quite flat, then see if you can improve the SNR by moving the microphone closer to the speaker's mouth, by removing noise sources in the environment, or by increasing the signal volume (for example, if you have a preamp which you can use to crank up the level of the signal before it is added to the noise inside the sound card, thus overpowering the noise and producing a high signal-to-noise ratio). If you can't get close to a flat noise floor in a waveform display under any circumstances, then you should definitely try a better mic or mic-plus-preamp setup.

Some recommended microphones

The Shure SM10A noise-cancelling headset microphone plus the Shure FP11 preamp. This mic/preamp setup is good for relatively noisy environments because not only does it have clean recording electronics, but also it improves the SNR both by keeping the mic 1-2cm from the mouth so that the speech drowns out everything else, and also by doing a certain amount of directional noise cancellation. This is the standard data collection mic setup in the speech R&D community, so most speech recognition systems have been developed using data that went through one of these, but the cost is on the order of $400(US) including the preamp. Other preamps I've seen used by speech people are available from ART and Rane. Preamp and mic cost around $110-150 for each one, and I got them from a professional music recording equipment store (Guitar Center, which has stores in at least San Jose and Seattle)

The Knowles VR-3565 is a good quality microphone which they have standardized on at Entropic in Cambridge England, where the speech recognition core technology development at Entropic is done. It costs in the neighborhood of $75.

The Andrea ANC-200 handset with preamp. This works great in trade shows because you can hand someone a handset and they talk to it like a phone, the sound is good. But it seems to be going out of production, and it seems to break easily. Cost was around $70.

Copyright © 1998-2003 Sprex, Inc. All rights reserved.
Modified: March 4, 2003