Speech Recognition Neural Network with Dynamic Frame Skipping

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition systems face increased calculation time due to the inability to perform parallel calculations as the number of frames in a speech signal increases, particularly when using neural networks as acoustic models.

Innovation Solution

The proposed method dynamically skips frames in the acoustic sequence based on predicted pronunciation duration using a neural network with a skip layer, allowing for efficient processing by minimizing information loss and reducing processing time, thereby enhancing processing speed and response time while maintaining recognition accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the number of frames in a speech signal increases to improve recognition accuracy, then the recognition accuracy is improved, but the calculation time increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidcalculation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The acoustic sequence is divided into multiple frames, and the neural network processes these segmented frames through multiple layers. This segmentation allows the system to handle large numbers of frames in a structured manner, improving recognition accuracy while managing calculation time through efficient layer-by-layer processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent employs a dynamic neural network architecture where the number of frames processed can be adjusted. The system dynamically determines the optimal number of frames to process based on the specific speech signal characteristics, allowing it to maintain high recognition accuracy while adapting calculation time to the actual processing needs rather than always processing maximum frames.

Inventive Principle:
Principle #15Dynamics

2Measurement precision

If the number of frames processed is increased to improve recognition accuracy, then the recognition accuracy is improved, but the processing speed decreases

Engineering Contradiction:
Improverecognition accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

By segmenting the acoustic sequence into frames and processing them through multiple neural network layers, the system achieves high recognition accuracy. The segmentation structure enables efficient parallel processing of frame features while maintaining the ability to handle large frame counts, thus preserving processing speed.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The neural network creates intermediate representations and feature copies at different layers, allowing the system to process frame information more efficiently. These copied representations enable the network to maintain high accuracy without reprocessing the same information multiple times, thereby improving processing speed.

Inventive Principle:
Principle #26Copying

3Measurement precision

If the number of frames processed is increased to improve recognition accuracy, then the recognition accuracy is improved, but the server costs increase

Engineering Contradiction:
Improverecognition accuracyVSAvoidserver costs
Core Design Contradiction:
Measurement precisionVSLoss of energy

Solution Approach 1:

The frame segmentation approach allows the system to process speech signals in manageable units through the neural network. This segmentation enables efficient resource utilization by processing only the necessary number of frames required for accurate recognition, reducing unnecessary computational overhead and server costs.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system changes processing parameters dynamically, adjusting the number of frames processed based on the specific speech signal characteristics and recognition requirements. This parameter adjustment allows the system to maintain high recognition accuracy while optimizing computational resource usage and reducing server costs by avoiding processing of excess frames.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3474274B1Speech recognition method and apparatus
Publication Date: 2021.07.14 SAMSUNG ELECTRONICS CO LTD
  • EP3474274B1 patent drawingFigure 1
  • EP3474274B1 patent drawingFigure 2
  • EP3474274B1 patent drawingFigure 3

AI summary

A speech recognition method includes obtaining an acoustic sequence, predicting a duration of a same pronunciation in the acoustic sequence, and skipping a pronunciation prediction for the duration.