Speech Recognition Using Phonetic Symbol Sequences for Unfamiliar Words

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition technologies face challenges in accurately recognizing words that are not fully learned or trained during the training process, such as proper nouns and newly-coined words, leading to reduced recognition accuracy.

Innovation Solution

The implementation of a speech recognition system that uses an encoder-decoder based model to convert speech input into text, where phonetic symbol sequences are used to represent unrecognized words, and a post-processing step that replaces these sequences with corresponding words using dictionary data and language models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a neural network-based speech recognition model is used, then generalization ability is improved, but recognition accuracy for unfamiliar words deteriorates

Engineering Contradiction:
Improvegeneralization abilityVSAvoidrecognition accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent introduces phonetic symbol sequences as an intermediary representation between speech input and text output. When the neural network encounters unfamiliar words during inference, it can output phonetic symbol sequences instead of direct text, which are then converted to words using a pronunciation dictionary. This intermediary mechanism allows the model to maintain generalization ability while improving accuracy for unfamiliar words.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent segments the speech recognition process into two distinct stages: (1) speech-to-phonetic-symbol conversion using the neural network, and (2) phonetic-symbol-to-word conversion using a pronunciation dictionary. This segmentation allows each component to specialize - the neural network handles general speech patterns while the dictionary handles exact word matching, resolving the contradiction between generalization and precision.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If phonetic symbol sequences are used to represent unrecognized words, then recognition accuracy for unfamiliar words is improved, but device complexity increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent makes the phonetic symbol sequence representation universal by using it for all types of unfamiliar words (proper nouns, newly-coined words, etc.) rather than creating separate handling mechanisms for each type. The same phonetic symbol sequence format and dictionary-based conversion process handles diverse word types, improving accuracy without proportionally increasing complexity.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent performs preliminary preparation by building a pronunciation dictionary that maps phonetic symbol sequences to words before the actual speech recognition process. This pre-computed resource enables fast, accurate conversion during inference without adding complex real-time processing, resolving the complexity-accuracy trade-off.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If dictionary data and language models are used for post-processing, then recognition accuracy is improved, but processing time increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies partial post-processing by using the pronunciation dictionary only for phonetic symbol sequences that the neural network outputs during inference, rather than processing the entire speech recognition pipeline. This selective application improves accuracy for unfamiliar words while minimizing additional processing time, as most common words are directly recognized by the neural network without dictionary lookup.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3772734B1Speech recognition method and apparatus
Publication Date: 2025.02.19 SAMSUNG ELECTRONICS CO LTD
  • EP3772734B1 patent drawingFigure 1
  • EP3772734B1 patent drawingFigure 2
  • EP3772734B1 patent drawingFigure 3

AI summary

A speech recognition method includes receiving speech data, obtaining, from the received speech data, a candidate text including at least one word and a phonetic symbol sequence associated with a pronunciation of a target word included in the received speech data, using a speech recognition model, replacing the phonetic symbol sequence included in the candidate text with a replacement word corresponding to the phonetic symbol sequence, and determining a target text corresponding to the received speech data based on a result of the replacing.