Acoustic Length Perturbation for DNN Generalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural network (DNN) acoustic models suffer from overfitting due to a large number of parameters, leading to poor generalization on new data, which can be addressed through length perturbation techniques that include frame skipping and frame insertion to alter the length of speech feature sequences.
Innovation Solution
Implementing frame skipping and frame insertion as data augmentation methods to perturb the length of acoustic utterances, which can be applied individually or in combination with other techniques like n-best based label smoothing to improve generalization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If DNN acoustic models use a large number of parameters to capture complex speech patterns, then modeling accuracy improves, but overfitting increases leading to poor generalization on new data
Solution Approach 1:
The patent applies length perturbation as a preliminary data augmentation technique during the training phase. By randomly inserting or deleting frames in speech sequences before training, the model is exposed to varied input lengths beforehand, which prevents overfitting and improves generalization performance on new data without sacrificing modeling accuracy
Solution Approach 2:
The patent changes the length parameter of speech sequences by inserting or deleting frames. This parameter modification creates diverse training samples with different sequence lengths, forcing the model to learn robust representations that generalize better to unseen data while maintaining the ability to capture complex speech patterns
2Productivity
If frame skipping is applied to reduce sequence length for faster processing, then training efficiency improves, but information loss increases affecting model accuracy
Solution Approach 1:
The patent applies frame skipping selectively and partially rather than uniformly. By randomly skipping only certain frames with a controlled probability, the method achieves faster processing and reduced computational load while preserving sufficient speech information to maintain model accuracy. The partial application ensures that not all critical information is lost
3Measurement precision
If frame insertion is applied to increase sequence length for better temporal representation, then temporal modeling improves, but computational complexity increases
Solution Approach 1:
The patent performs frame insertion as a preliminary data augmentation step during training data preparation. By pre-generating training samples with inserted frames, the model learns better temporal representations without incurring high computational complexity during actual inference. The computational burden is shifted to the offline training phase where it can be managed more efficiently
Data Source
AI summary
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.