Subword-Based Speech Recognition Network Structure

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition systems face limitations in speed due to character-based end-to-end models and complexity with word-based models, and subword-based models suffer from training-recognition mismatches.

Innovation Solution

The implementation of a subword-based end-to-end speech recognition apparatus that uses a subword dictionary to simplify the network structure, estimate sequences of subwords, and convert them into recognizable words, minimizing mismatches between training and recognition operations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If a character-based end-to-end acoustic model is used, then the model can process speech signals, but the output length becomes too long and recognition speed is limited

Engineering Contradiction:
Improverecognition speedVSAvoidoutput length
Core Design Contradiction:
SpeedVSDevice complexity

Solution Approach 1:

The patent segments words into subwords (e.g., morphemes, syllables, or character n-grams) to create a intermediate representation. This segmentation reduces the output length compared to character-based models while maintaining better linguistic structure than word-based models. The subword units serve as building blocks that can be efficiently processed by the acoustic model.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different processing strategies to different parts of the speech recognition pipeline. The acoustic model uses subword-based output for local efficiency, while a separate language model component handles global linguistic constraints. This local optimization allows speed improvement in the acoustic processing stage without sacrificing overall recognition accuracy.

Inventive Principle:
Principle #3Local quality

2Ease of manufacture

If a word-based end-to-end acoustic model is used, then recognition can be performed, but the great number of output classes makes implementation difficult

Engineering Contradiction:
Improveimplementation easeVSAvoidnumber of output classes
Core Design Contradiction:
Ease of manufactureVSDevice complexity

Solution Approach 1:

The patent divides the large vocabulary into smaller subword units, reducing the number of output classes the acoustic model must handle directly. Instead of outputting thousands of word classes, the model outputs a smaller set of subword classes that can be recombined to form complete words, making implementation more manageable.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension in the output space by using subword representations. Rather than directly mapping acoustic features to word classes (one-dimensional mapping), the system maps to subword sequences that can be composed into words, effectively adding a compositional dimension to the recognition process.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Productivity

If a subword-based end-to-end acoustic model is used, then output length is reduced, but mismatches occur between training and recognition due to unrecognized subwords

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidtraining-recognition consistency
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent performs preliminary preparation by creating a comprehensive subword dictionary during the training phase that anticipates all possible subword units that may appear during recognition. This preliminary action ensures that the model is trained on the same subword vocabulary it will encounter during recognition, preventing mismatches.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent implements feedback mechanisms where the system monitors for unrecognized subwords during recognition and uses this information to improve the subword dictionary and model. This feedback loop ensures continuous alignment between training and recognition capabilities, reducing mismatches over time.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3624114B1Method and apparatus for speech recognition
Publication Date: 2021.08.11 SAMSUNG ELECTRONICS CO LTD
  • EP3624114B1 patent drawingFigure 1
  • EP3624114B1 patent drawingFigure 2
  • EP3624114B1 patent drawingFigure 3

AI summary

A processor-implemented speech recognition method includes: extracting a speech feature from an input speech to be recognized; estimating a first sequence of first subwords corresponding to at least one portion of the input speech based on the extracted speech feature; converting the first sequence to a second sequence of at least one second subword by combining at least two of the first subwords; and recognizing the input speech by recognizing a remaining portion of the input speech based on the second sequence.