Adaptive Decoder Confusion Reduction for Compressed ASR

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Speech recognition systems on edge devices face challenges due to hardware constraints such as limited memory and power, leading to increased error rates when using compressed acoustic models, and confusability issues between similar commands.

Innovation Solution

A system that dynamically expands the speech lexicon by recovering grapheme or phoneme sequences with high error rates, determines confusability metrics, and replaces confusable commands with semantically equivalent but phonetically distinct alternate commands, allowing user input to select replacements, thereby reducing confusability and improving accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If a compressed acoustic model is used on edge devices, then memory constraints are satisfied, but error rates increase

Engineering Contradiction:

Improvememory usageVSAvoiderror rate

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

The system dynamically adjusts the complexity parameter of the acoustic model based on available memory resources. It transitions between compressed and uncompressed models, and selectively loads different components (acoustic model, language model, lexicon) to optimize the balance between memory consumption and recognition accuracy.

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The speech recognition system implements dynamic model loading and unloading mechanisms. The acoustic model can be compressed during low-memory conditions and decompressed when memory is available, allowing the system to adapt its resource consumption in real-time based on device state and operational requirements.

Inventive Principle:

Principle #15Dynamics

2Quantity of substance

If model compression is applied to meet space requirements, then memory constraints are satisfied, but classification accuracy deteriorates

Engineering Contradiction:

Improvemodel sizeVSAvoidclassification accuracy

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The speech recognition system is divided into separable components: acoustic model, language model, and lexicon. This segmentation allows selective compression and loading of individual components based on memory availability, enabling the system to maintain higher accuracy by loading full-precision models for specific tasks while keeping overall model size manageable through selective component loading.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system introduces an intermediary layer (language model and lexicon) that compensates for the reduced accuracy of compressed acoustic models. The language model provides contextual disambiguation and the lexicon offers precise word matching, thereby recovering classification accuracy that is lost during acoustic model compression.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Measurement precision

If the speech lexicon is expanded to recover grapheme or phoneme sequences, then recognition accuracy is improved, but confusability between similar commands increases

Engineering Contradiction:

Improverecognition accuracyVSAvoidcommand distinction

Core Design Contradiction:

Measurement precisionVSReliability

Solution Approach 1:

The system implements a feedback mechanism that monitors confusability metrics between command pairs. When confusability exceeds a threshold, the system automatically adjusts the lexicon by removing or replacing one of the confusable commands, thereby maintaining command distinction while preserving overall recognition accuracy through adaptive lexicon management.

Inventive Principle:

Principle #23Feedback

4Reliability

If confusable commands are replaced with phonetically distinct alternates, then command distinction is improved, but system adaptability to user preferences is reduced

Engineering Contradiction:

Improvecommand distinctionVSAvoiduser preference accommodation

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system dynamically manages the lexicon based on real-time confusability analysis. When confusable commands are detected, the system can selectively replace only the more problematic commands while preserving user-preferred alternatives, allowing adaptive adjustment between command distinction and user preference accommodation based on operational context.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS12067978B2Methods and systems for confusion reduction for compressed acoustic models

Publication Date: 2024.08.20 SAMSUNG ELECTRONICS CO LTD

AI summary

Methods and systems are disclosed herein for improvements relating to compressed automatic speech recognition (ASR) systems. The ASR system may comprise a compressed acoustic engine and an adaptive decoder. The adaptive decoder may be dynamically compiled based on characteristics of the compressed acoustic engine and a current state of the application device. In some embodiments, a dynamic command list is used to manage context-specific commands. Two or more commands recognized by the adaptive decoder may be confusable due to compression of the ASR system. Alternate commands may be determined that are semantically equivalent but phonetically different than the confusable commands to reduce classification error of the adaptive decoder. An alternate command may replace one or more of the confusable commands in the adaptive decoder. In some embodiments, a user interface is displayed to a user of the ASR system to select the alternate command for replacement in the decoder.