Singular value decomposition for improved voice recognition in presence of multi-talker background noise
View Patent ↗A system and method for providing speech recognition functionality offers improved accuracy and robustness in noisy environments having multiple speakers. The described technique includes receiving speech energy and converting the received speech energy to a digitized form. The digitized speech energy is decomposed into features that are then projected into a feature space having multiple speaker subspaces. The projected features fall either into one of the multiple speaker subspaces or outside of all speaker subspaces. A speech recognition operation is performed on a selected one of the multiple speaker subspaces to resolve the utterance to a command or data.
1. A method of providing speech recognition functionality for an utterance spoken in a noisy environment having multiple human speakers, the method comprising:
receiving speech energy corresponding to the utterance, the received speech energy comprising contributions from multiple human speakers;
converting the received speech energy to an electronic form;
digitizing the electronic form of the received speech energy to render a digitization of the received speech energy;
decomposing the digitization of the received speech energy to produce feature data representative of features in the digitization of the received speech energy, wherein the feature data does not distinguish between different speakers and between speech and background noise;
processing the feature data to produce speaker dependent feature data and speaker independent feature data;
projecting only the speaker independent feature data into a feature space, the feature space having multiple human speaker subspaces for multiple human speakers, wherein each one of the multiple human speakers is associated with a distinct one of the multiple human speaker subspaces and wherein the features of the speaker independent feature data project either into one of the multiple human speaker subspaces or outside of all human speaker subspaces;
identifying the speaker independent feature data associated with a speaker subspace associated with a primary human speaker; and
performing a speech recognition operation on speaker independent feature data associated with the speaker subspace associated with the primary human speaker to resolve the utterance to a command or data,
wherein the speaker independent feature data includes contributions from at least two speakers, and
wherein performing a speech recognition operation on the feature data associated with the speaker subspace associated with the primary human speaker to resolve the utterance to a command or data comprises removing all feature data not associated with the speaker subspace associated with the primary human speaker.
2. The method of providing speech recognition functionality according to claim 1 , wherein the step of converting the received speech energy to an electronic form comprises transducing audible speech energy to an analog electronic signal.
3. The method of providing speech recognition functionality according to claim 1 , wherein the step of processing the feature data to produce speaker dependent feature data and speaker independent feature data includes applying a singular value decomposition matrix transformation to the data.
4. The method of providing speech recognition functionality according to claim 1 , wherein the utterance is a command, and wherein performing a speech recognition operation on the speaker independent feature data associated with the speaker subspace associated with the primary human speaker to resolve the utterance to a command or data comprises resolving the utterance to a command, and wherein the method further comprises the subsequent step of causing the command to be executed.
5. The method of providing speech recognition functionality according to claim 1 , wherein each human speaker subspace contains only speaker related phoneme information.
6. The method of providing speech recognition functionality according to claim 5 , wherein the speaker related phoneme information is speaker specific normalized data.
7. The method of providing speech recognition functionality according to claim 1 , wherein the step of projecting only the speaker independent feature data into the feature space comprises projecting the speaker independent feature data into the multiple human speaker subspaces by using eigenvalues and orthonormal bases.
8. A method of providing speech recognition functionality for an utterance spoken in a noisy environment having multiple human speakers, the method comprising:
extracting feature data from the utterance, wherein the feature data comprises one or more features;
processing the feature data to produce speaker dependent feature data and speaker independent feature data;
projecting only the speaker independent data into a feature space, the feature space having multiple human speaker subspaces for multiple human speakers, wherein each one of the multiple human speakers is associated with a distinct one of the multiple human speaker subspaces;
identifying speaker independent feature data associated with a speaker subspace associated with a primary human speaker; and
performing a speech recognition operation restricted to a single one of the multiple human speaker subspaces,
wherein the speaker independent feature data includes contributions from at least two speakers, and
wherein performing a speech recognition operation restricted to a single one of the multiple human speaker subspaces comprises removing all feature data not associated with the speaker subspace associated with the primary human speaker.
9. The method of providing speech recognition functionality based on a spoken utterance according to claim 8 , wherein the utterance comprises speech energy and wherein extracting feature data from the utterance comprises:
receiving the speech energy, converting the received speech energy to an electronic form, digitizing the converted speech energy, and decomposing the digitized speech energy into features.
10. The method of claim 8 , further comprising:
identifying speaker independent feature data associated with a second speaker subspace associated with a second human speaker; and
performing a second speech recognition operation restricted to a second one of the multiple human speaker subspaces.
11. The method of providing speech recognition functionality based on a spoken utterance according to claim 10 , further comprising dividing the input speech energy by contribution from each of the at least two speakers.
12. The method of providing speech recognition functionality based on a spoken utterance according to claim 8 , wherein the step of projecting the speaker independent feature data into a feature space comprises projecting the speaker independent feature data into the multiple human speaker subspaces by using eigenvalues and orthonormal bases.
13. The method of providing speech recognition functionality based on a spoken utterance according to claim 8 , wherein processing the feature data to produce speaker dependent feature data and speaker independent feature data includes applying a singular value decomposition matrix transformation to the feature data.
14. The method of providing speech recognition functionality based on a spoken utterance according to claim 8 , wherein the step of performing a speech recognition operation restricted to a single one of the multiple human speaker subspaces comprises discarding all data not in the selected one of the multiple speaker subspaces.
15. The method of providing speech recognition functionality based on a spoken utterance according to claim 8 , wherein the utterance is a command, and wherein performing a speech recognition operation restricted to a single one of the respective one or more speaker subspaces comprises resolving the utterance to a command, and wherein the method further comprises the subsequent step of causing the command to be executed.
16. The method of providing speech recognition functionality based on a spoken utterance according to claim 8 , wherein each of the multiple human speaker subspaces contains only speaker related phoneme information.
17. The method of providing speech recognition functionality based on a spoken utterance according to claim 16 , wherein the speaker related phoneme information is speaker specific normalized data.
18. A system for providing speech recognition functionality for an utterance spoken in a noisy environment having multiple human speakers, the system comprising:
a transducer for converting the utterance to an electrical speech signal;
a digitizer for digitizing the electrical speech signal;
a feature extraction module for extracting one or more features from the digitized electrical speech signal to produce feature data;
a processing module for processing the feature data to produce speaker dependent feature data and speaker independent feature data;
a projection module for projecting the speaker independent feature data into a feature space, the feature space having multiple human speaker subspaces for multiple human speakers, wherein each one of the multiple human speakers is associated with a distinct one of the multiple human speaker subspaces;
an identification engine configured to identify speaker independent feature data associated with a speaker subspace associated with a primary human speaker; and
a recognition engine is configured to perform a speech recognition operation restricted to a single one of the multiple human speaker subspaces to resolve the spoken utterance,
wherein the speaker independent feature data includes contributions from at least two speakers, and
wherein performing a speech recognition operation restricted to a single one of the multiple human speaker subspaces to resolve the spoken utterance comprises removing all feature data not associated with the speaker subspace associated with the primary human speaker.
19. The system for providing speech recognition functionality according to claim 18 , wherein the recognition engine is further adapted to cause a command to be executed if the utterance is resolved to the command.