METHOD AND SYSTEM FOR GENERATING ADVANCED FEATURE DISCRIMINATION VECTORS FOR USE IN SPEECH RECOGNITION

• US 20160284343A1
• Filed: 03/17/2014
• Published: 09/29/2016
• Est. Priority Date: 03/15/2013
• Status: Active Grant

• Current Assignee: XMOS Ltd.
• Original Assignee: XMOS Ltd.
• Inventors: Hone, Brian, Short, Kevin M.
• Granted Patent: US 9,728,182 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units
• CPC Class Codes: G10L 25/03 characterised by the type of extracted parameters
• CPC Class Codes: G10L 25/18 the extracted parameters being spectral information of each sub-band
• CPC Class Codes: G10L 25/21 the extracted parameters being power information
• CPC Class Codes: G10L 25/24 the extracted parameters being the cepstrum
• CPC Class Codes: G10L 25/93 Discriminating between voiced and unvoiced parts of speech signals G10L25/90 takes precedence

Abstract

A method of renormalizing high-resolution oscillator peaks, extracted from windowed samples of an audio signal, is disclosed. Feature vectors are generated for which variations in both fundamental frequency and time duration of speech are substantially mitigated. The feature vectors may be aligned within a common coordinate space, free of those variations in frequency and time duration that occurs between speakers, and even over speech by a single speaker, to facilitate a simple and accurate determination of matches between those AFDVs generated from a sample of the audio signal and corpus AFDVs generated for known speech at the phoneme and sub-phoneme level. The renormalized feature vectors can be combined with traditional feature vectors such as MFCCs, or they can be used exclusively to identify voiced, semi-voiced and unvoiced sounds.

Claims

1. A method of generating advanced feature discrimination vectors (AFDVs) representing sounds forming at least part of an input audio signal, the method comprising:

taking a plurality of samples of the input audio signal, each of the samples being a portion of the input audio signal as it evolves over a window of predetermined time;

for each sample of the audio signal taken;

performing a signal analysis on the sample to extract one or more high resolution oscillator peaks therefrom, the extracted oscillator peaks forming a spectral representation of the sample;

renormalizing the extracted oscillator peaks to eliminate variations in the fundamental frequency and time duration for each sample occurring over the window;

normalizing the power of the renormalized extracted oscillator peaks; and

forming the renormalized and power normalized extracted oscillator peaks into an AFDV for the sample.

2. The system of claim 1, wherein the renormalizing further comprises:

for each of the samples of the audio signal analyzed;

identifying a harmonically related group of the extracted oscillator peaks, each oscillator peak of the identified group being located in a frequency bin dictated by the spectral representation of the windowed sample of the audio signal; and

shifting the oscillator peaks to consecutive frequency bins beginning with bin 1, while maintaining the relative ordering of the identified oscillator peaks from the spectral representation of the windowed sample.

3. The method of claim 2 wherein the extracted renormalized oscillator peaks for each sample analyzed are distributed into slots of a comparator stack to form a common coordinate system by which to directly compare the AFDV for each sample to AFDVs representing known speech sounds stored in a library for purposes of facilitating the identification of a closest match therebetween.

4. The method of claim 3, wherein the comparator stack has n slots among which the oscillator peaks for the identified group are distributed evenly, the oscillator peak occupying a highest frequency bin number always being located in the n^th slot.

5. The method of claim 3 wherein the AFDVs for each analyzed sample is representative of a voiced sound.

6. The method of claim 2 wherein if no harmonically related group of oscillator peaks can be identified for an analyzed sample of the audio input signal, forming the extracted oscillator peaks into a feature vector for that sample that represents a weighted average of the spectral power over sub-bands of frequency covering the entire frequency range of the spectral representation.

7. The method of claim 6 wherein the feature vector that represents a weighted average of the spectral power over the sub-bands of frequency covering the entire frequency range of the spectral representation of the windowed sample is an MFCC.

8. The method of claim 3, wherein if no harmonically related group of oscillator peaks can be identified for an analyzed sample of the audio signal, distribute the extracted oscillator peaks into the comparator stack in accordance with a weighted average of the power for oscillator peaks falling within predetermined bands of frequencies.

9. The method of claim 8 wherein the AFDV for an analyzed sample of the audio signal having no identifiable harmonically related group of extracted oscillator peaks represents an unvoiced sound.

10. The method of claim 3 wherein oscillator peaks of significant power exist in addition to the identified group of harmonically related oscillator peaks, the method further comprising:

distributing the oscillator peaks of the identified group of harmonically related oscillator peaks evenly into slots forming a first subdivision of the comparator stack; and

distributing the additional oscillator peaks, not part of the identified group, into slots forming a second subdivision of the comparator stack in accordance with a weighted average of the power of the additional oscillator peaks falling within predetermined bands of frequencies.

11. The method of claim 10 wherein the oscillator peaks of the identified group are representative of a voiced component of a semi-voiced sound, and the additional peaks are representative of an unvoiced component of the semi-voiced sound.

12. The method of claim 11, wherein the power normalizing is performed in manner that maintains the ratio of the power between the peaks in the first subdivision and the second subdivision.

13. The method of claim 12, wherein the identification of a closest match is made based on a dot product of the AFDV for each sample to AFDVs representing the known speech sounds.

14. The method of claim 1, wherein the normalizing of power normalizes the overall power of the renormalized oscillator peaks to a value of 1.

15. The method of claim 1 wherein the signal analysis employs complex spectral phase evolution (CSPE) to extract the oscillator peaks.

16. The method of claim 1 wherein at least one of the extracted oscillator peaks is identified as short-time stable oscillator.

17. The method of claim 1 wherein at least one of the extracted oscillator peaks is frequency modulated over the time of window.

18. The method of claim 1 wherein at least one of the extracted oscillator peaks is amplitude modulated over the time of the window.

19. The method of claim 1 wherein the signal analysis includes eliminating the effect of the window in extracting the oscillator peaks

20. The method of claim 1 wherein a tracker is used to determine which of the extracted oscillator peaks are sounds emanating from a source of interest.