Speech Recognition Accuracy via Feature Vector Clustering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional speech recognition techniques require large amounts of training data to effectively improve speech recognition accuracy, especially in scenarios with limited data, such as short voice clips, leading to inefficiencies and reduced accuracy.

Innovation Solution

The method involves clustering feature vectors of training data to reduce their number and complexity, followed by interpolation calculations using clustered feature vectors during the recognition stage, which compensates for the accuracy of data to be recognized, particularly suitable for short voice clips.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large amount of training data is used to obtain feature vectors, then speech recognition accuracy is improved, but data requirement and processing complexity increase

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidtraining data volume
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent segments the training data by clustering feature vectors into multiple clusters, where each cluster represents a subset of the data with similar characteristics. This segmentation allows the system to work with smaller, more manageable data subsets while maintaining recognition accuracy, directly addressing the contradiction between accuracy and data volume requirements

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms the raw feature vectors into clustered feature vectors, changing the parameter representation from individual data points to cluster centroids. This parameter transformation reduces the effective data volume needed for training while preserving the essential patterns needed for accurate speech recognition

Inventive Principle:
Principle #35Parameter changes

2Productivity

If feature vectors are clustered to reduce data volume, then processing efficiency is improved, but information loss may occur

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidfeature vector information
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent creates clustered feature vectors as representative copies of groups of original feature vectors. Each cluster centroid serves as a compressed representation that captures the essential characteristics of multiple data points, maintaining information fidelity while reducing processing load

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent performs clustering as a preliminary processing step before the main speech recognition task. By pre-organizing the feature vectors into clusters and computing cluster centroids in advance, the system prepares compressed representations that speed up subsequent processing without losing critical information

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3479377B1Speech recognition
Publication Date: 2023.12.06 ALIBABA GROUP HOLDING LTD
  • EP3479377B1 patent drawingFigure 1
  • EP3479377B1 patent drawingFigure 2
  • EP3479377B1 patent drawingFigure 3

AI summary

A speech recognition method includes clustering feature vectors of training data to obtain clustered feature vectors of training data performing interpolation calculation on feature vectors of data to be recognized using the clustered feature vectors of training data, and inputting the feature vectors of data to be recognized after the interpolation calculation into a speech recognition model to adaptively adjust the speech recognition model. The techniques of the present disclosure improve speech recognition accuracy and adaptive processing efficiency.