Context-Bound Speech Recognition With Domain-Aware Language Mixing

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional domain-specific ASR systems are highly inaccurate due to the lack of sufficient training data and the inability to leverage out-of-domain language, leading to reduced accuracy in toxic speech detection and other domain-specific applications.

Innovation Solution

A context-aware domain-specific language model is trained using smart context injections, filtering general dataset sentences or phrases with domain words, generating n-grams, and interpolating with a general language model to enhance accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional domain-specific ASR language models are used for toxic speech detection, then the system can detect offensive and inappropriate speech, but the detection accuracy is highly inaccurate

Engineering Contradiction:

Improvedetection accuracyVSAvoidmodel accuracy

Core Design Contradiction:

Measurement precisionVSReliability

Solution Approach 1:

The patent combines a domain-specific language model with a general language model through interpolation. The domain-specific model is trained on toxic speech data while the general model provides broader language context, merging their strengths to achieve both specialized detection accuracy and overall reliability

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The language model system uses composite training data consisting of both domain-specific toxic speech corpora and general language corpora. This composite approach creates a more robust model that maintains accuracy across different speech types while specifically improving toxic speech detection

Inventive Principle:

Principle #40Composite materials

2Productivity

If domain-specific training data is limited, then the model can be trained faster with less data, but the recognition accuracy deteriorates

Engineering Contradiction:

Improvetraining efficiencyVSAvoidrecognition accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The general language model component provides universal language understanding that complements the domain-specific model. This multi-functional approach allows the system to achieve high recognition accuracy without requiring extensive domain-specific training data, as the general model fills gaps where domain data is limited

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Adaptability or versatility

If the language model is trained only on domain-specific data, then the model becomes specialized for toxic speech detection, but it cannot leverage out-of-domain language effectively

Engineering Contradiction:

Improvedomain specializationVSAvoidout-of-domain language understanding

Core Design Contradiction:

Adaptability or versatilityVSLoss of information

Solution Approach 1:

The general language model acts as an intermediary that bridges domain-specific toxic speech detection with broader language understanding. It mediates between the specialized domain model and general language patterns, preventing information loss about out-of-domain language while maintaining domain specialization

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS12555572B2Method and system of automatic context-bound domain-specific speech recognition

Publication Date: 2026.02.17 INTEL CORP

AI summary

A system, article, and method of automatic context-bound domain-specific speech recognition uses general language models.