(JAN 2009) — The release of LumenVox 8.6 — now available for Windows and Linux — features a number of new features, but probably the most exciting is our new noise reduction technology.
This new technology helps the Speech Engine better deal with callers in noisy environments — think users calling from inside their cars while speeding down a highway.
The noise reduction works by sampling the audio passed into the Speech Engine before the caller starts speaking. This is about the first half-second of audio while a prompt is playing. The Engine examines that background audio and determines a sort of acoustic signature for it, a mathematical representation of the background noise.
Once the Engine has the signature of the background noise, it is able to subtract that signature from the entire audio stream — effectively removing the background noise without damaging the actual signal that is the user's speech.
In developing this technology, we did quite a bit of testing to ensure that this process does not interfere with clean audio. One danger in doing noise reduction is that if it mistakes signal for noise, it can remove important information, damaging recognition quality.
To confirm that our new process does not harm good clean audio, we performed a battery of tests with the Engine, using clean and noisy data, comparing the accuracy with noise reduction and the accuracy without the reduction.
The chart below displays the results of these tests. This chart shows the average accuracy of the Engine at recognizing digits for various levels of noise. The "clean" column represents audio with almost no noise while the other columns represent the signal:noise ratio. This ratio is how much louder the signal (the user's speech) is than the background noise. At 0dB, the background noise is at exactly the same volume as the speech.
|Without noise reduction||94.5||87.0||80.4||68.3||54.9||38.2||70.55|
|With noise reduction||94.4||93.3||86.0||69.6||55.8||54.8||75.65|
Note: The above chart is for noisy data and does not represent LumenVox accuracy in most situations.
As you can see, the Engine with noise reduction performs signifcantly better at almost all levels of noise. It performs only slightly worse with clean data, introducing an extra error only 1 in 1,000 times.
We feel this extremely small error rate is a good trade off for the overall improved accuracy — on average, we receive 5 more correct answers per 100 digits with the noise reduction.
To take advantage of this new technology, all you need to do is to upgrade to LumenVox 8.6. This technology is built–in to the 8.6 speech client, so if you upgrade your systems to the latest release you'll be able to begin using it today.
In your LumenVox.com Account, you can now pay for your software maintenance online. This means that if you're not current on your annual software maintenance, you can renew it online and get immediate access to the new software.
In addition to the noise reduction technology, LumenVox 8.6 includes a number of other new changes.
Some new configuration files now exist, allowing you greater runtime flexibility with the Speech Engine and any applications that make use of the LumenVox API.
You can alter many settings via these configuration files instead of using the API calls, which means you can tweak Engine options without changing your code. And the old methods still work, meaning there should be total backwards compatibility with older applications.
We've also added some new routines to our servers and our client libraries related to startup, which should help reduce errors and improve the ease of troubleshooting problems that do occur.
For more details on any of these changes, see the latest Release Notes in our help file.
© 2016 LumenVox, LLC. All rights reserved.