Length perturbation techniques for improving generalization of deep neural network acoustic models
One or more systems, devices, computer program products and/or computer-implemented methods of use provided herein relate to length perturbation techniques for improving generalization of DNN acoustic models. A computer-implemented system can comprise a memory that can store computer executable components. The computer-implemented system can further comprise a processor that can execute the computer executable components stored in the memory, wherein the computer executable components can comprise a frame skipping component that can remove one or more frames from an acoustic utterance via frame skipping. The computer executable components can further comprise a frame insertion component that can insert one or more replacement frames into the acoustic utterance via frame insertion to replace the one or more frames with the one or more replacement frames to enable length perturbation of the acoustic utterance.
1 . A computer-implemented system, comprising:
a memory that stores computer executable components; and
a processor that executes at least one of the computer executable components that:
performs length perturbation of an acoustic utterance comprising a group of frames in a sequence, wherein the performing the length perturbation comprises:
random sampling a first defined percentage of frames from the group of frames resulting in a first subset of drop frames;
for each drop frame, removing a first defined quantity of consecutive frames from the group of frames starting with the drop frame;
random sampling a second defined percentage of frames from the group of frames resulting in a second subset of insert frames; and
for each insert frame, inserting a second defined quantity of replacement frames into the acoustic utterance after the insert frame.
2 . The computer-implemented system of claim 1 , wherein the at least one of the computer executable components further:
receives the acoustic utterance from a speech recognition system.
3 . The computer-implemented system of claim 1 , wherein the length perturbation of the acoustic utterance assists with generalization of a deep neural network (DNN) acoustic model.
4 . The computer-implemented system of claim 1 , wherein the length perturbation of the acoustic utterance is a one-way perturbation enabled by at least one of the removing the first defined quantity of consecutive frames or the inserting of the second defined quantity of replacement frames.
5 . The computer-implemented system of claim 1 , wherein the length perturbation of the acoustic utterance is a two-way perturbation enabled by both the removing the first defined quantity of consecutive frames and the frame inserting of the second defined quantity of replacement frames.
6 . The computer-implemented system of claim 1 , wherein at least one of the computer executable components applies the length perturbation of the acoustic utterance as individual data augmentation technique.
7 . The computer-implemented system of claim 1 , wherein at least one of the computer executable components applies the length perturbation of the acoustic utterance as a first data augmentation technique in combination with at least a second data augmentation technique that is different from the length perturbation.
8 . The computer-implemented system of claim 1 , wherein the length perturbation of the acoustic utterance is applied to one or more domains of feature representations of the acoustic utterance.
9 . The computer-implemented system of claim 1 , wherein the the removing the first defined quantity of consecutive frames and the frame inserting of the second defined quantity of replacement frames are applied towards the length perturbation using one or more sequence-to-sequence models, to assist with generalization of the one or more sequence-to-sequence models.
10 . The computer-implemented system of claim 1 , wherein the the removing the first defined quantity of consecutive frames and the frame inserting of the second defined quantity of replacement frames are applied towards the length perturbation using one or more sequence-to-one models, to assist with generalization of the one or more sequence-to-one models.
11 . A computer-implemented method, comprising:
performing, by a system operatively coupled to a processor, length perturbation of an acoustic utterance comprising a group of frames in a sequence, wherein the performing the length perturbation comprises:
random sampling a first defined percentage of frames from the group of frames resulting in a first subset of drop frames;
for each drop frame, removing a first defined quantity of consecutive frames from the group of frames starting with the drop frame;
random sampling a second defined percentage of frames from the group of frames resulting in a second subset of insert frames; and
for each insert frame, inserting a second defined quantity of replacement frames into the acoustic utterance after the insert frame.
12 . The computer-implemented method of claim 11 , further comprising:
receiving, by the system, the acoustic utterance from a speech recognition system.
13 . The computer-implemented method of claim 11 , wherein the length perturbation of the acoustic utterance assists with generalization of a DNN acoustic model.
14 . The computer-implemented method of claim 11 , wherein the length perturbation of the acoustic utterance is a one-way perturbation enabled by at least one of the removing the first defined quantity of consecutive frames or the inserting of the second defined quantity of replacement frames.
15 . The computer-implemented method of claim 11 , wherein the length perturbation of the acoustic utterance is a two-way perturbation enabled by both the removing the first defined quantity of consecutive frames and the frame inserting of the second defined quantity of replacement frames.
16 . The computer-implemented method of claim 11 , further comprising:
applying, by the system, the length perturbation of the acoustic utterance as an individual data augmentation technique.
17 . The computer-implemented method of claim 11 , further comprising:
applying, by the system, the length perturbation of the acoustic utterance as a first data augmentation technique in combination with at least a second data augmentation technique that is different from the length perturbation.
18 . The computer-implemented method of claim 11 , further comprising:
applying, by the system, the length perturbation of the acoustic utterance to one or more domains of feature representations of the acoustic utterance.
19 . A computer program product for improving generalization of a DNN acoustic model, the computer program product comprising a non-transitory computer readable medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:
performing, by the processor, length perturbation of an acoustic utterance comprising a group of frames in a sequence, wherein the performing the length perturbation comprises:
random sampling a first defined percentage of frames from the group of frames resulting in a first subset of drop frames;
for each drop frame, removing a first defined quantity of consecutive frames from the group of frames starting with the drop frame;
random sampling a second defined percentage of frames from the group of frames resulting in a second subset of insert frames; and
for each insert frame, inserting a second defined quantity of replacement frames into the acoustic utterance after the insert frame.
20 . The computer program product of claim 19 , wherein the program instructions are further executable by the processor to cause the processor to:
receive, by the processor, the acoustic utterance from a speech recognition system.