Multi-source phoneme classification for noise-robust automatic speech recognition
View Patent ↗A system and method are disclosed for processing an audio signal including separating the audio signal into a plurality of streams which group sounds from a same source prior to classification and analyzing each separate stream to determine phoneme-level classification. One or more words of the audio signal may then be outputted.
1. A method of processing an audio signal comprising:
computing 600 spectral values on a logarithmic frequency scale from the audio signal;
separating the 600 spectral values into a plurality of streams which group sounds from a same source prior to classification;
analyzing each separated stream to determine phoneme-level classification; and
outputting one or more words of the audio signal.
2. The method of processing an audio signal as recited in claim 1 wherein phoneme-level classification accuracy is enhanced by providing as an input to a classifier a spectral envelope.
3. The method of processing an audio signal as recited in claim 1 wherein phoneme-level classification accuracy is enhanced by providing as an input to a classifier detected transients.
4. The method of processing an audio signal as recited in claim 1 wherein phoneme-level classification accuracy is enhanced by providing as an input to a classifier pitch and voicing information.
5. The method of processing an audio signal as recited in claim 1 further comprising normalizing for speaker characteristics prior to classification.
6. The method of processing an audio signal as recited in claim 1 further comprising performing noise-threshold tracking.
7. The method of processing an audio signal as recited in claim 6 further comprising adjusting gain and setting noise floor reference levels based on the noise-threshold tracking.
8. The method of processing an audio signal as recited in claim 1 further comprising training with a full phoneme target set.
9. The method of processing an audio signal as recited in claim 1 further comprising incorporating a model of syllabic stress.
10. The method of processing an audio signal as recited in claim 1 wherein the spectral values are computed with a 6 microsecond resolution post-interpolation.
11. The method of processing an audio signal as recited in claim 1 wherein the spectral values are updated every 22 microseconds.
12. The method of processing an audio signal as recited in claim 1 further comprising using the output for automatic voice-dialing for phones.
13. The method of processing an audio signal as recited in claim 1 further comprising using the output for automatic command of a system or device.
14. The method of processing an audio signal as recited in claim 1 further comprising using the output as an interface to a device.
15. The method of processing an audio signal as recited in claim 1 wherein the output comprises a meeting transcription.