Speech Recognition Using Syllable Substitution for Low-Frequency Words

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition technologies face challenges in accurately recognizing a wide range of words due to memory size limitations in phonetic dictionaries, leading to inefficiencies in word sequence generation and recognition accuracy.

Innovation Solution

A method and apparatus that generate a word sequence based on a phoneme sequence, with a syllable sequence substitution for words with lower recognition rates, using both word unit and syllable unit phonetic and language models to enhance recognition accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large number and diversity of words are registered in the phonetic dictionary to enhance speech recognition accuracy, then recognition accuracy improves, but memory size requirements increase

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidmemory size
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The phonetic dictionary is segmented into multiple parts: a first phonetic dictionary stored in memory containing frequently used words, and a second phonetic dictionary stored externally containing less frequently used words. This segmentation allows the system to maintain high recognition accuracy for common words while managing memory constraints by offloading less critical data to external storage.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system transitions from a single-dimensional memory storage approach to a multi-dimensional architecture involving both internal memory and external storage. By distributing the phonetic dictionary across different storage dimensions (memory vs. external storage), the system achieves both fast access for common words and comprehensive coverage for rare words without being constrained by limited memory size.

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

2Adaptability or versatility

If every possible word combination is registered in the phonetic dictionary to improve recognition accuracy, then coverage of diverse words improves, but device complexity increases

Engineering Contradiction:
Improveword coverageVSAvoidphonetic dictionary management complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

Different parts of the phonetic dictionary are assigned different qualities based on their usage frequency. The first phonetic dictionary in memory contains high-priority, frequently used words optimized for fast access, while the second phonetic dictionary in external storage contains low-priority, less frequently used words. This local quality differentiation allows the system to manage complexity by treating different word subsets differently based on their importance.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system dynamically selects which phonetic dictionary to use based on the recognition context and frequency requirements. For common words, it quickly accesses the first phonetic dictionary in memory, while for less common words, it accesses the second phonetic dictionary in external storage. This dynamic selection mechanism allows the system to adapt to different recognition scenarios without maintaining all words in high-speed memory.

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If words with lower recognition rates are handled using traditional methods, then processing speed is maintained, but recognition accuracy deteriorates

Engineering Contradiction:
Improverecognition accuracy for low-frequency wordsVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary classification of words into frequent and less frequent categories before recognition processing. By pre-organizing the phonetic dictionary into two separate dictionaries based on usage frequency, the system can quickly determine which dictionary to access during recognition, avoiding time-consuming searches through the entire dictionary while ensuring accurate recognition for both common and rare words.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentUS10140974B2Method and apparatus for speech recognition
Publication Date: 2018.11.27 SAMSUNG ELECTRONICS CO LTD
  • US10140974B2 patent drawing
  • US10140974B2 patent drawing
  • US10140974B2 patent drawing

AI summary

A method and apparatus for speech recognition are disclosed. The speech recognition apparatus includes a processor configured to process a received speech signal, generate a word sequence based on a phoneme sequence generated from the speech signal, generate a syllable sequence corresponding to a word element among words comprised in the word sequence based on the phoneme sequence, and determine a text corresponding to a recognition result of the speech signal based on the word sequence and the syllable sequence.