Speech Recognition Candidate Selection Using Inter-Sentence Distance

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing speech recognition techniques fail to consider the relevance between sentences, leading to inaccuracies in associating multiple sentences from an audio signal.

Innovation Solution

A speech recognition apparatus that generates candidate text groups for each audio segment and selects the most relevant text pairs across segments based on inter-sentence distances, enhancing the accuracy of sentence recognition.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If chain probability between successive words is considered for sentence acquisition, then word-level recognition is improved, but sentence-level relevance is not considered

Engineering Contradiction:

Improveword recognition accuracyVSAvoidsentence relevance

Core Design Contradiction:

Measurement precisionVSReliability

Solution Approach 1:

The patent segments the speech recognition task into two independent stages: first generating candidate sentences based on word chain probability, then selecting the optimal sentence combination based on inter-sentence distance. This segmentation allows each stage to optimize for its specific objective without compromising the other.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension of evaluation by calculating inter-sentence distance between candidate sentences. This adds a sentence-level relevance dimension to the traditional word-level chain probability approach, enabling simultaneous optimization for both word accuracy and sentence coherence.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

2Adaptability or versatility

If multiple candidate sentences are generated from audio data, then recognition coverage is improved, but selection accuracy decreases

Engineering Contradiction:

Improverecognition coverageVSAvoidsentence selection accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the selecting unit evaluates candidate sentence combinations using inter-sentence distance calculations and selects the optimal combination. This feedback loop ensures that the final selected sentences are not only generated from the audio data but also validated for their coherence and relevance.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent changes the selection parameter from simple word-level chain probability to a composite metric that includes inter-sentence distance. This parameter transformation enables the system to distinguish between grammatically correct but contextually irrelevant sentences and those that are both grammatically correct and contextually appropriate.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12586585B2Speech recognition apparatus, control method, and non-transitory storage medium

Publication Date: 2026.03.24 NEC CORP

AI summary

A speech recognition apparatus (2000) acquires a plurality of pieces of audio data (20) for a source audio signal including an utterance. The speech recognition apparatus (2000) generates a candidate text group (30) for each of the plurality of pieces of audio data (20). The candidate text group (30) includes a plurality of candidate texts (32). The candidate text (32) is a candidate of a text representing a content of an utterance corresponding to the audio data (20), and represents a sentence. The speech recognition apparatus (2000) selects, based on a comparison result between the plurality of candidate text groups (30), for each of the pieces of audio data (20), a candidate text (32) representing a content of an utterance represented by the piece of audio data (20) from the candidate text group (30) generated for the piece of audio data (20).