Systems and methods for accelerating automatic speech recognition based on compression and decompression
Systems and methods are provided for automatic speech recognition. In the method, the system obtains a padded sequence by processing a plurality of acoustic signals. The system compresses the padded sequence by reducing the size of the padded sequence to obtain a compressed sequence. The system inputs the compressed sequence into a pre-trained encoder neural network to obtain an encoded sequence and then decompresses the encoded sequence by recovering the encoded sequence to an original sequential ordering. The system inputs the encoded sequence to a decoding module to obtain recognition texts.
1. A method for automatic speech recognition, comprising:
obtaining a padded sequence by processing a plurality of acoustic signals associate with a speech input;
compressing the padded sequence by reducing a size of the padded sequence to obtain a compressed sequence;
inputting the compressed sequence into a pre-trained encoder neural network to obtain an encoded sequence;
decompressing the encoded sequence by recovering the encoded sequence to an original sequential ordering to obtain a decompressed sequence prior to an inter-frames computation performed within the pre-trained encoder neural network, wherein the inter-frames computation takes place between frames; and
inputting the encoded sequence that is decompressed to a decoding module and generating recognition texts as an output from the decoding module.
2. The method of claim 1 , wherein reducing the size of the padded sequence comprises removing previously added zero frames from the padded sequence.
3. The method of claim 2 , wherein recovering the encoded sequence to the original sequential ordering comprises restoring removed zero frames to the encoded sequence.
4. The method of claim 2 , further comprising:
generating an offset mask matrix based on positional information of the removed zero frames.
5. The method of claim 4 , wherein recovered the encoded sequence to the original sequential ordering comprises restoring the removed zero frames to the encoded sequence according to the offset mask matrix.
6. The method of claim 4 , wherein the positional information comprises a plurality of prefixed sums calculated based on the locations of the removed zero frames within the padded sequence.
7. The method of claim 1 , further comprising:
compressing the decompressed sequences by reducing the size of the decompressed sequence prior to an intra-frame computation performed within the pre-trained encoder network, wherein the intra-frame computation takes place within the frames.
8. The method of claim 2 , further comprising:
evaluating an efficiency gain prior to an intra-frame computation performed within the pre-trained encoder network, wherein the intra-frame computation takes place within the frames; and
compressing the encoding sequences based on the evaluated efficiency gain by removing padded empty frames.
9. The method of claim 8 , wherein the efficiency gain is evaluated based on a throughput value, a query-per-second (QPS) value, or a latency value.
10. An automatic speech recognition system for recognizing speech, comprising:
at least one computer storage memory operable to store data along with computer-executable instructions; and
at least one processor operable to read the data and operate the computer-executable instructions to:
obtaining a padded sequence by processing a plurality of acoustic signals associate with a speech input;
compressing the padded sequence by reducing a size of the padded sequence to obtain a compressed sequence;
inputting the compressed sequence into a pre-trained encoder neural network to obtain an encoded sequence;
decompressing the encoded sequence by recovering the encoded sequence to an original sequential ordering to obtain a decompressed sequence prior to an inter-frames computation performed within the pre-trained encoder neural network, wherein the inter-frames computation takes place between frames; and
inputting the encoded sequence that is decompressed to a decoding module and generating recognition texts as an output from the decoding module.
11. The automatic speech recognition system of claim 10 , wherein reducing the size of the padded sequence comprises removing previously added zero frames from the padded sequence.
12. The automatic speech recognition system of claim 11 , wherein recovering the encoded sequence to the original sequential ordering comprises restoring removed zero frames to the encoded sequence.
13. The automatic speech recognition system of claim 11 , wherein the at least one processor is further configured to:
generating an offset mask matrix based on positional information of the removed zero frames.
14. The automatic speech recognition system of claim 13 , wherein recovered the encoded sequence to the original sequential ordering comprises restoring the removed zero frames to the encoded sequence according to the offset mask matrix.
15. The automatic speech recognition system of claim 13 , wherein the positional information comprises a plurality of prefixed sums calculated based on the positions of the removed zero frames within the padded sequence.
16. The automatic speech recognition system of claim 10 , wherein the at least one processor is further configured to:
compressing the decompressed sequences by reducing the size of the encoded sequence prior to an intra-frame computation performed within the pre-trained encoder network, wherein the intra-frame computation takes place within the frames.
17. The automatic speech recognition system of claim 11 , wherein the at least one processor is further configured to:
evaluating an efficiency gain prior to an intra-frame computation performed within the pre-trained encoder network, wherein the intra-frame computation takes place within the frames; and
compressing the encoding sequences based on the evaluated efficiency gain by removing padded empty frames.
18. The automatic speech recognition system of claim 17 , wherein the efficiency gain is evaluated based on a throughput value, a query-per-second (QPS) value, or a latency value.
19. A non-transitory computer readable medium having stored thereon a program for performing a method of automatic speech recognition, the method comprising:
obtaining, by at least one processor, a padded sequence by processing a plurality of acoustic signals associate with a speech input;
compressing, by the at least one processor, the padded sequence by reducing a size of the padded sequence to obtain a compressed sequence;
encoding, by a pre-trained encoder neural network by the at least one processor, the compressed sequence to obtain an encoded sequence;
decompressing, by the at least one processor, the encoded sequence by recovering the encoded sequence to an original sequential ordering to obtain a decompressed sequence prior to an inter-frames computation performed within the pre-trained encoder neural network, wherein the inter-frames computation takes place between frames; and
decoding, by a decoding module in the at least one processor, the encoded sequence that is decompressed; and
generating, by the decoding module, recognition texts as an output of the decoding module.