Subword Segmentation Regularization for OOV Speech Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Modern automated speech recognition (ASR) systems face challenges in handling out-of-vocabulary (OOV) terms and learning relationships between linguistically-related words using word-based representations, while grapheme-based representations require larger memory footprints and computational times, leading to non-optimal performance.

Innovation Solution

Implementing a subword segmentation routine that segments input words into subword units using a trained vocabulary set, applying regularization techniques like misspelling and character swapping to improve ASR model performance, and using architectures like RNN-T or LAS for efficient speech recognition.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If Chinese characters are displayed without word segmentation, then the display is simple and continuous, but the meaning cannot be properly conveyed and reading comprehension is impaired

Engineering Contradiction:
Improvemeaning conveyance accuracyVSAvoidword segmentation complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent introduces a word segmentation layer as an intermediary between the Chinese character display layer and the meaning interpretation layer. This segmentation layer divides continuous Chinese characters into meaningful word units using segmentation rules and data structures, enabling accurate meaning conveyance without requiring changes to the fundamental continuous display nature of Chinese characters.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies segmentation by dividing the continuous stream of Chinese characters into discrete word units based on linguistic rules and semantic boundaries. Each word unit is represented as a structured data element containing position information, character data, and semantic meaning, allowing the system to process and interpret Chinese text meaningfully while maintaining visual continuity.

Inventive Principle:
Principle #1Segmentation

2Reliability

If word segmentation is implemented to improve reading comprehension, then meaning conveyance is enhanced, but the processing complexity and computational resources increase

Engineering Contradiction:
Improvereading comprehension accuracyVSAvoidprocessing system complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent changes the parameter representation of Chinese text from a simple continuous character stream to a structured array of word units with multiple attributes (position, characters, meaning, length). This parameter transformation enables sophisticated processing and interpretation while providing a standardized framework that manages computational complexity through consistent data structures.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces manual or rule-based word segmentation mechanisms with an automated computing system that uses algorithms and data structures to perform segmentation. This substitution reduces manual intervention complexity while systematically handling the computational tasks required for accurate word boundary detection and meaning assignment.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

3Reliability

If traditional word segmentation methods are used, then some meaning can be conveyed, but irregularities and inconsistencies in segmentation results occur

Engineering Contradiction:
Improvesegmentation consistencyVSAvoidsegmentation rule complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies homogeneity by establishing uniform segmentation rules and consistent data structures for all Chinese text processing. The system uses standardized criteria for identifying word boundaries, assigning meanings, and structuring word units, ensuring that the same input always produces the same output format regardless of the specific text content, thereby achieving high segmentation consistency.

Inventive Principle:
Principle #33Homogeneity

Solution Approach 2:

The patent incorporates feedback mechanisms where segmentation results are validated and refined based on linguistic rules and semantic consistency checks. The system adjusts segmentation decisions by evaluating contextual information and applying correction rules, ensuring that irregular cases are resolved consistently and that segmentation accuracy improves through iterative refinement.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4305544B1Regularizing word segmentation
Publication Date: 2026.05.06 GOOGLE LLC
  • EP4305544B1 patent drawingFigure 1
  • EP4305544B1 patent drawingFigure 2A
  • EP4305544B1 patent drawingFigure 2B

AI summary

A method (600) for subword, segmentation includes receiving an input word (302) to be segmented into a plurality of subword units (119). The method also includes executing a subword segmentation routine (300) to segment the input word into a plurality of subword units by accessing a trained vocabulary set (350) of subword units and selecting the plurality of subword units from the input word by greedily finding a longest subword unit from the input word that is present in the trained vocabulary set until an end of the input word is reached.