If you read the patent application, noise suppression is only part of the picture. In earlier versions of speech recognition the computer would isolate the spectral pattern of the voice and use that pattern for analyses. The problem was that when a person moved their head their spectral pattern would change.
The approach in this patent uses a phoneme database at the front-end of a speech recognition engine connected to the monitor (Acoustic Model Selector ACM). The ACM selects sounds (after noise suppression is performed ) which best match regular speech sounds (vowels and consonants). They are moved to a back-end process which match the phoneme to a language and then works out the command.
The combination of multiple mics and software algorithms help in 'beam formation' of the voice ie selecting the best signal to noise ratio, which is then used to eliminate noise. The characteristics of the set-up (mic placement, type, axis etc) are stored in software and can be used by the speech recognition software to account for spectral characteristics. There also seem to be a few combinations of best mic placement as well as a 4 mic option depending on the display characteristics.
Beam formation is different to phase inversion as a noise reduction technique. My version on how this works is that the best signal from both highly directional mics are matched to form a beam, while other signals, which are weaker are suppressed.
A phase inversion technique requires an omni directional mic to monitor background noise and another directional mic for the voice (Or a polar set-up). The background noise would be phase inverted and added to the voice signal to remove the noise. This is really only a good technique if the noise is ambient ie, all around you. Eg reflected factory noise. I'm not sure this patent makes use of this technique.
The approach in this patent uses a phoneme database at the front-end of a speech recognition engine connected to the monitor (Acoustic Model Selector ACM). The ACM selects sounds (after noise suppression is performed ) which best match regular speech sounds (vowels and consonants). They are moved to a back-end process which match the phoneme to a language and then works out the command.
The combination of multiple mics and software algorithms help in 'beam formation' of the voice ie selecting the best signal to noise ratio, which is then used to eliminate noise. The characteristics of the set-up (mic placement, type, axis etc) are stored in software and can be used by the speech recognition software to account for spectral characteristics. There also seem to be a few combinations of best mic placement as well as a 4 mic option depending on the display characteristics.
Beam formation is different to phase inversion as a noise reduction technique. My version on how this works is that the best signal from both highly directional mics are matched to form a beam, while other signals, which are weaker are suppressed.
A phase inversion technique requires an omni directional mic to monitor background noise and another directional mic for the voice (Or a polar set-up). The background noise would be phase inverted and added to the voice signal to remove the noise. This is really only a good technique if the noise is ambient ie, all around you. Eg reflected factory noise. I'm not sure this patent makes use of this technique.
