Finite State Transducer Speech Recognition Combining Language Models

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Consumer-grade speech recognition systems lack domain-specific vocabulary and phraseology, leading to high error rates for professionals, making it economically inefficient for them to use these systems, as they require specialized terminology that is not captured by general-purpose recognition systems.

Innovation Solution

The technology combines Finite State Transducers (FSTs) to customize speech recognition systems by integrating general-purpose and user-specific language models in real-time, allowing for on-demand recognition of domain-specific vocabulary and phraseology without the need to create new decoders for each user, using independent segments or back-off approaches to facilitate rapid combination of language models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If consumer-grade speech recognition systems are used, then the system works acceptably for a large number of users, but the accuracy is insufficient for professionals who require domain-specific vocabulary and phraseology

Engineering Contradiction:

Improveadaptability to domain-specific vocabularyVSAvoidrecognition accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The language model is segmented into multiple components: a general-purpose language model for common vocabulary and domain-specific language models for specialized terminology. The system selectively activates appropriate segments based on the input context, allowing high accuracy for both general and domain-specific speech recognition without requiring a single massive model

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The speech recognition system is designed to be universal by incorporating multiple language models that can handle different domains. The system automatically selects or combines appropriate language models based on the input, making it adaptable to both consumer-grade and professional domain-specific requirements with a single unified system

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Measurement precision

If a custom speech recognition system with domain-specific vocabulary is created for each professional user, then recognition accuracy improves, but computational time and storage requirements increase significantly

Engineering Contradiction:

Improverecognition accuracyVSAvoidcomputational time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

Language models are pre-computed and stored in an optimized format during system setup. When a professional user needs domain-specific recognition, the pre-prepared language model is quickly loaded and combined with the acoustic model, avoiding the need to compute custom models in real-time and reducing both storage and computational time

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system merges pre-computed language models with acoustic models in an optimized manner. By combining these models efficiently during the recognition process rather than creating entirely new custom systems, the patent reduces computational time while maintaining high recognition accuracy for domain-specific vocabulary

Inventive Principle:

Principle #5Merging (Combining)

3Measurement precision

If a custom speech recognition system with domain-specific vocabulary is created for each professional user, then recognition accuracy improves, but storage requirements increase significantly

Engineering Contradiction:

Improverecognition accuracyVSAvoidstorage requirements

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The language model is segmented into reusable components that can be shared across multiple users. Domain-specific vocabulary and phraseology are organized into modular language model segments that can be selectively combined with general-purpose models, reducing redundant storage while maintaining high recognition accuracy for each user's specific domain

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system uses universal language model components that serve multiple purposes and users. By designing language models that can be flexibly combined and reused across different professional domains, the patent reduces overall storage requirements while still providing customized high-accuracy recognition for each user's specific vocabulary and phraseology

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS9966066B1System and methods for combining finite state transducer based speech recognizers

Publication Date: 2018.05.08 NVOQ INC

AI summary

A speech recognition engine is provided. The speech recognition engine combines multiple, different statistical language models to allow a single instance of a speech recognition engine to recognize words applicable to a first group of users and a second group of users where the second group of users is a subset of the first group of users. At least one of the multiple, different statistical language models comprises words specific to the second group of users.