NOISE CANCELLATION: Reducing Noise With Software
As speech technology continues to mature, noise is emerging as a key limiter of future performance. This magazine has often addressed noise canceling microphone hardware. Here we review the corresponding approach of noise removal processing in software. Software and hardware noise reduction are complementary. Directional microphones reduce both continuous and discrete noise events from "off-axis" locations. Software processing reduces continuous noise from all sources and directions, including any internal noise from the microphone and sound card circuitry. The figure below shows noisy speech recorded simultaneously with different noise canceling microphones, versus final waveforms after software noise reduction. Both microphones passed appreciable noise, which software processing was able to remove. Input speech distortion is another concern. Here, each microphone gave quite different results from identical inputs, while the software processing caused no added distortions in either case.
Speech-to-Text Dictation requires very low word error rate (WER) for wide acceptance. Even with a headset microphone and in a "quiet" office, the "hum" of background noise limits achievable error rates for large vocabulary dictation systems. This is shown from testing the VoiceType dictation system, both with and without software noise processing. Tests were done in an office, with no background speakers, using an Andrea ANC-600 headset microphone. Two different recordings were made, one for enrollment, another for testing. The speech-to-noise ratios (SNRs) were 30-40 dB. One copy of VoiceType was trained and tested with no software noise reduction, and another was trained and tested with the software noise processing. Training and testing without the software noise reduction gave 76 errors in 1009 words of spoken text. Training and testing with added software filtering gave 22 errors out of the same 1009 words of spoken text.
| | No Filtering | Filtering |
WER in Quiet Office | 7.5% | 2.2% |
Specialized voice command systems are more noise resistant than large vocabulary dictation systems. Yet, they often must work in higher noise levels for automotive, industrial, and military applications. Enrollment and training data were recorded and mixed with noise levels from a military vehicle environment. Enrollment and testing were done both with and without software noise reduction. For voice command applications, a false positive response is a critical failure – worse than no response at all. Results below summarize both correct responses, and also false positives.
| No Noise Filtering | With Noise Filtering |
SNR (dB) | Correct | False Positives | Correct | False Positives |
35 | 100% | 0% | 100% | 0% |
20 | 88% | 4% | 99% | 1% |
10 | 6% | 4% | 77% | 0% |
Speaker verification systems often have noisy inputs, from background noise, telephone lines, laptop microphones, and the like. False rejection of the valid speaker can result from either failure to detect the speech or failure to recognize the speaker. Testing results below are for a commercial speaker verification system, with and without software noise reduction included. Enrollment was in a quiet environment (30 dB or higher SNR), either with or without software noise reduction. Corresponding tests were done over a range of SNRs. Results show that without software noise reduction the system becomes unusable at any SNR below 27 dB. With software noise reduction useful performance is maintained down to SNRs of 15 dB or less.
More information about software noise reduction, including audio demonstrations, can be found at
,www.ca.defgrp.com/n_prefilter.html and at other links from that site. Morgan Grover and David Makovoz work for Defense Group Incorporated. They can be reached at: Defense Group Incorporated, (310) 394-8599,
morgan@ca.defgrp.com,
www.ca.defgrp.com.