2 ms·

The noise suppression demo is incredible! Scroll down to "Show Me the Samples!" (about 3/4 of the way down) on https://people.xiph.org/~jm/demo/rnnoise/ https:/

The noise suppression demo is incredible! Scroll down to "Show Me the Samples!" (about 3/4 of the way down) on https://people.xiph.org/~jm/demo/rnnoise/ https://people.xiph.org/~jm/demo/rnnoise/ and hit play and try the buttons. My heavens; this sort of technology would enable, say, a robot to have a much better time understanding what a person is saying in a public place (as well as having possibilities for better chatting for people over the internet).

jononor 8y ago
One can do as you propose for speech recognition, use noise reduction first, then a classifier. But it can be more efficient to just train a speech recognition detector directly on input corrupted by noise. This way the detector learns invariant to noise. This is commonly done via data augmentation, which mixes in various noise into labeled speech snippets. There are still[1] circumstances where one would have noise suppression as a pre-processing step to the speech detector. For instance when using microphone arrays, the frontend may use data from many microphones (2-20), perform adaptive Beamforming extract the cleanest possible mono speech signal. This can include a Vocal Activity Detection (VAD) as estimator, and multi-source aware noise reduction. The output is then ran through a standard mono detector, either on device (typically only Keyword Spotting) or sending to the cloud. Strong neural networks in combination with microphone arrays is the reason why smart home devices like Alexa etc have become pretty decent (compared to what was feasible 10 years ago) at speech recognition when the speaker is far away. 1. Integrating more and more into functions into big neural networks and jointly optimizing the overall system is definitely a trend, and very actively researched.

2 ms·

The noise suppression demo is incredible! Scroll down to "Show Me the Samples!" (about 3/4 of the way down) on https://people.xiph.org/~jm/demo/rnnoise/ https:/

by armagon 8y ago

jononor 8y ago
One can do as you propose for speech recognition, use noise reduction first, then a classifier. But it can be more efficient to just train a speech recognition detector directly on input corrupted by noise. This way the detector learns invariant to noise. This is commonly done via data augmentation, which mixes in various noise into labeled speech snippets. There are still[1] circumstances where one would have noise suppression as a pre-processing step to the speech detector. For instance when using microphone arrays, the frontend may use data from many microphones (2-20), perform adaptive Beamforming extract the cleanest possible mono speech signal. This can include a Vocal Activity Detection (VAD) as estimator, and multi-source aware noise reduction. The output is then ran through a standard mono detector, either on device (typically only Keyword Spotting) or sending to the cloud. Strong neural networks in combination with microphone arrays is the reason why smart home devices like Alexa etc have become pretty decent (compared to what was feasible 10 years ago) at speech recognition when the speaker is far away. 1. Integrating more and more into functions into big neural networks and jointly optimizing the overall system is definitely a trend, and very actively researched.