Method and system for speech recognition
Embodiments of the disclosure provide a method and system for speech recognition. The method comprises dividing space into a plurality of regions based on preset DOA angles to allocate a signal source to the plurality of regions, wherein signals in the plurality of regions are enhanced and recognized, the result of which are fused to obtain a recognition result of the signal source.
1. A method comprising:
allocating a signal source based on different directions of arrival (DOAs) by dividing a physical space into a plurality of non-overlapping regions to allocate the signal source into the plurality of regions, the allocation performed prior to beamforming, the plurality of non-overlapping regions based on preset DOA angles comprising at least two of: an angle of 30 degrees, an angle of 60 degrees, an angle of 90 degrees, an angle of 120 degrees, and an angle of 150 degrees;
enhancing signals of the signal source for each of the regions to obtain enhanced signals corresponding to the regions, the enhancing performed independently for each region;
performing speech recognition on the enhanced signals corresponding to the regions to obtain recognition results corresponding to the regions;
providing the recognition results corresponding to the regions to respective acoustic models, each acoustic model trained for its corresponding region based on enhanced signal samples from that region; and
fusing outputs of the acoustic models to obtain a recognition result, wherein fusing analyzes outputs from all preset regions regardless of estimated signal source direction.
2. The method of claim 1 , the enhancing signals of the signal source corresponding to the regions comprising performing delay-and-sum (DAS) beamforming on the signals of the signal source corresponding to the regions to obtain the enhanced signals.
3. The method of claim 1 , the enhancing signals of the signal source corresponding to the regions comprising performing Minimum Variance Distortionless Response (MVDR) beamforming on the signals of the signal source corresponding to the regions to obtain the enhanced signals.
4. The method of claim 1 , further comprising:
dividing, prior to the allocating the signal source, a space into regions according to the regions;
performing speech enhancement on speech signals in the different regions to obtain different enhanced signal samples; and
using the obtained samples to perform training to obtain the acoustic models corresponding to the regions.
5. The method of claim 1 , the fusing output results is performed by using a Recognizer Output Voting Error Reduction (ROVER) based fusion system.
6. A system comprising:
a processor; and
a storage medium for tangibly storing thereon program logic for execution by the processor, the stored program logic comprising:
logic, executed by the processor, for allocating a signal source based on different directions of arrival (DOAs) by dividing a physical space into a plurality of non-overlapping regions to allocate the signal source into the plurality of regions, the allocation performed prior to beamforming, the plurality of non-overlapping regions based on preset DOA angles comprising at least two of: an angle of 30 degrees, an angle of 60 degrees, an angle of 90 degrees, an angle of 120 degrees, and an angle of 150 degrees,
logic, executed by the processor, for enhancing signals of the signal source for each of the regions to obtain enhanced signals corresponding to the regions, the enhancing performed independently for each region, respectively,
logic, executed by the processor, for performing speech recognition on the enhanced signals corresponding to the regions to obtain recognition results corresponding to the regions, each acoustic model trained for its corresponding region, logic, executed by the processor, for providing the recognition results corresponding to the regions to respective acoustic models based on enhanced signal samples from that region, and
logic, executed by the processor, for fusing output results from the acoustic models to obtain a recognition result, wherein the fusing analyzes outputs from all preset regions regardless of estimated signal source direction.
7. The system of claim 6 , the logic for allocating a signal source based on regions comprising: logic, executed by the processor, for dividing a space into a plurality of regions to allocate the signal source into the plurality of regions formed based on DOA angles.
8. The system of claim 6 , the logic for enhancing signals of the signal source corresponding to the regions comprising: logic, executed by the processor, for performing delay-and-sum (DAS) beamforming on the signals of the signal source corresponding to the regions to obtain the enhanced signals.
9. The system of claim 6 , the logic for enhancing signals of the signal source corresponding to the regions comprising: logic, executed by the processor, for performing Minimum Variance Distortionless Response (MVDR) beamforming on the signals of the signal source corresponding to the regions to obtain the enhanced signals.
10. The system of claim 6 , the stored program logic further comprising:
logic, executed by the processor, prior to the allocating the signal source, for dividing a space into regions according to the regions,
logic, executed by the processor, for performing speech enhancement on speech signals in the different regions to obtain different enhanced signal samples, and
logic, executed by the processor, for using the obtained samples to perform training to obtain the acoustic models corresponding to the regions.
11. The system of claim 6 , the fusing output results is performed by using a Recognizer Output Voting Error Reduction (ROVER) based fusion system.
12. A non-transitory computer-readable storage medium for tangibly storing computer program instructions capable of being executed by a computer processor, the computer program instructions defining:
allocating a signal source based on different directions of arrival (DOAs) by dividing a physical space into a plurality of generally non-overlapping regions to allocate the signal source into the plurality of regions, the allocation performed prior to beamforming, the plurality of non-overlapping regions based on preset DOA angles comprising at least two of: an angle of 30 degrees, an angle of 60 degrees, an angle of 90 degrees, an angle of 120 degrees, and an angle of 150 degrees;
enhancing signals of the signal source for each of the regions to obtain enhanced signals corresponding to the regions, the enhancing performed independently for each region;
performing speech recognition on the enhanced signals corresponding to the regions to obtain recognition results corresponding to the regions;
providing the recognition results corresponding to the regions to respective acoustic models, each acoustic model trained for its corresponding region based on enhanced signal samples from that region; and
fusing output results from the acoustic models to obtain a recognition result, wherein the fusing analyzes outputs from all preset regions regardless of estimated signal source direction.
13. The computer-readable storage medium of claim 12 , the allocating a signal source based on regions comprising: dividing a space into a plurality of regions to allocate the signal source into the plurality of regions.
14. The computer-readable storage medium of claim 12 , the enhancing signals of the signal source corresponding to the regions comprising: performing delay-and-sum (DAS) beamforming on the signals of the signal source corresponding to the regions to obtain the enhanced signals.
15. The computer-readable storage medium of claim 12 , the enhancing signals of the signal source corresponding to the regions comprising performing Minimum Variance Distortionless Response (MVDR) beamforming on the signals of the signal source corresponding to the regions to obtain the enhanced signals.
16. The computer-readable storage medium of claim 12 , further comprising:
dividing, prior to the allocating the signal source, a space into regions according to the regions;
performing speech enhancement on speech signals in the different regions to obtain different enhanced signal samples; and
using the obtained samples to perform training to obtain the acoustic models corresponding to the regions.