Language Processing Device for Speech Recognition Error Correction

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Neural network models struggle to accurately process speech recognition errors in tasks like conversation summarization, as they do not account for phonological connections that result in errors with different meanings, leading to inaccuracies in interpreting speech recognition results.

Innovation Solution

A language processing device that generates error sentences based on pronunciation, using a neural network model to predict and update language model parameters, thereby accounting for phonological connections and improving accuracy in speech recognition errors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If speech recognition is used to convert speech utterances into text data, then language processing tasks such as conversation summarization can be performed, but phonological errors occur where words with similar pronunciation but different meanings are mistakenly recognized

Engineering Contradiction:

Improvelanguage processing capabilityVSAvoidaccuracy of text interpretation

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The system performs pre-training with phonological error sentences before actual language processing tasks. Error sentences are generated in advance by replacing tokens with phonologically similar alternatives, and the language model is trained to correct these errors, preparing the model to handle speech recognition errors in advance

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

Instead of directly training the model to process correct sentences, the system inverts the approach by training with erroneous sentences as input and correct sentences as target. The model learns to transform error-containing sentences into correct ones, thereby gaining robustness against phonological errors

Inventive Principle:

Principle #13The other way round (Inversion)

2Measurement precision

If standard word filling tasks are used for pre-training, then the language model can be trained, but phonological connections and speech recognition error tendencies are not considered

Engineering Contradiction:

Improvelanguage model training accuracyVSAvoidability to handle speech recognition errors

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The system applies local quality by specifically addressing phonological error patterns in certain positions within sentences. Error sentences are generated by replacing tokens with phonologically similar alternatives at specific positions, and the model learns to correct these localized errors while maintaining overall sentence structure

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The system changes the training parameter by introducing phonological similarity as a factor in error generation. Instead of random token replacement, errors are generated based on phonological characteristics, changing the nature of the training data to better reflect real speech recognition error patterns

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20250021762A1Language processing apparatus, image processing method, and program

Publication Date: 2025.01.16 NIPPON TELEGRAPH & TELEPHONE CORP

AI summary

A language processing device includes circuitry configured to generate an error sentence corresponding to an original sentence based on pronunciation corresponding to text data indicating the original sentence; use a language model based on a neural network model to generate a prediction sentence from the error sentence based on a language model parameter of the language model; and update the language model parameter based on a difference between the original sentence and the prediction sentence.