Named-Entity Voice Labeling for Domain-Specific STT Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing real-time Speech-to-Text (STT) services face challenges in accurately recognizing proper nouns and entities due to conflicts with commonly used terms and the need for extensive training, especially in specialized domains like finance and e-commerce, where proper nouns are frequently used and quickly changing.
Innovation Solution
A labeling method for uttered voice that automatically labels training data by extracting named entities from customer utterances, correcting errors through a user interface, and using domain-specific STT models to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a universal STT model is used for real-time speech-to-text conversion, then the service can be provided broadly across different domains, but the recognition accuracy for domain-specific proper nouns and entities deteriorates due to conflicts with commonly used terms
Solution Approach 1:
The patent divides the STT service into two segments: a universal STT model for general speech-to-text conversion and a domain-specific proper noun recognition model for accurate entity identification. This segmentation allows the system to maintain broad service coverage while achieving high accuracy for domain-specific terms by processing different types of recognition tasks through specialized models
Solution Approach 2:
The patent introduces an intermediary component that bridges the universal STT model and domain-specific proper nouns. This intermediary uses pre-trained proper noun models and contextual analysis to resolve ambiguities between common terms and domain-specific entities, enabling the universal model to accurately recognize proper nouns without requiring domain-specific retraining
2Measurement precision
If extensive training is conducted on domain-specific proper nouns to improve recognition accuracy, then the STT model performance for those entities improves, but the training time, labor, cost, and effort increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training proper noun recognition models using curated domain-specific datasets before deployment. This pre-training establishes a foundation of domain knowledge that can be quickly applied during real-time STT operations without requiring extensive on-demand training, thereby reducing training time and effort while maintaining high recognition accuracy
Solution Approach 2:
The patent changes the training parameters by using transfer learning and fine-tuning techniques. Instead of training from scratch with extensive data, the system leverages pre-trained language models and adapts them to domain-specific proper nouns through targeted fine-tuning with smaller, curated datasets, significantly reducing training time and computational resources
3Adaptability or versatility
If new products and buzzwords are continuously introduced, then the service remains relevant and up-to-date, but the difficulty of quickly training on newly used proper nouns increases due to the need for supervised learning
Solution Approach 1:
The patent implements a dynamic proper noun recognition system that can adapt to new products and buzzwords in real-time. The system uses continuous learning mechanisms and automated entity extraction from domain-specific sources to dynamically update its proper noun vocabulary without requiring complex retraining processes, enabling the service to remain current with minimal operational complexity
Solution Approach 2:
The patent enables the STT system to self-update its proper noun recognition capabilities by automatically extracting new entities from domain-specific data sources and incorporating them into the recognition model. This self-service mechanism eliminates the need for manual data collection and model retraining, allowing the system to adapt to new products and buzzwords autonomously
Data Source
AI summary
A labeling method for an uttered voice, performed by a computing system, comprises receiving a first uttered voice from a user terminal, acquiring a first uttered text by converting the first uttered voice into text, extracting a named entity included in the first uttered text by performing Named Entity Recognition (NER) on the first uttered text, acquiring, from a call agent terminal connected via a voice communication session with the user terminal, a second uttered voice including a pronunciation of a corrected named entity corresponding to the extracted named entity, and labeling the corrected named entity in the second uttered voice.