Speech Recognition Acoustic Embedding Noise Robustness

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current speech recognition technologies face challenges in accurately recognizing speech in noisy environments and efficiently classifying patterns, as they struggle to replicate the human brain's pattern recognition capabilities effectively.

Innovation Solution

The method involves converting words in speech data into continuous vector values based on phonetic similarities using an acoustic embedding matrix, allowing phonetically similar words to be positioned adjacent to each other in a vector space, and utilizing a two-stage language model approach for recognition, where candidate results are generated and refined using a first and second language model respectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional speech recognition methods are used, then the system can process speech data, but the recognition accuracy deteriorates in noisy environments

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidnoise interference
Core Design Contradiction:
Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent introduces acoustic embedding vectors as an intermediary representation between raw speech data and the language model. These embeddings transform speech inputs into a continuous vector space where phonetically similar words are positioned adjacently, serving as a mediator that filters noise while preserving phonetic information for improved recognition accuracy

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transforms discrete word representations into continuous vector parameters through acoustic embedding. By changing the representation parameter from discrete tokens to continuous vectors with phonetic similarity properties, the system achieves robustness against noise while maintaining recognition precision

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If a single language model is used for speech recognition, then the system complexity is reduced, but the recognition accuracy in challenging conditions deteriorates

Engineering Contradiction:
Improverecognition accuracyVSAvoidlanguage model structure
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the language model into two distinct components: a first language model that generates candidate recognition results and a second language model that refines these candidates. This segmentation allows each model to specialize in different aspects of recognition, improving overall accuracy while distributing computational complexity across multiple focused components

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The first language model performs preliminary action by generating candidate recognition results before the final decision is made. This preliminary processing step filters and ranks potential outcomes, allowing the second language model to focus computational resources on refining the most promising candidates, thereby improving accuracy without proportionally increasing overall complexity

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3154054B1Method and apparatus for recognizing speech
Publication Date: 2020.12.23 SAMSUNG ELECTRONICS CO LTD
  • EP3154054B1 patent drawingFigure 1
  • EP3154054B1 patent drawingFigure 2
  • EP3154054B1 patent drawingFigure 3

AI summary

A method and apparatus for training a neural network language model, and a method and apparatus for recognizing speech data based on a trained language model are provided. The method of training a language model involves converting, using a processor, training data into error-containing training data, and training a neural network language model using the error-containing training data.