Support Vector Machine Model Shrinking for Embedded Keyword Spotting

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current speech recognition systems require significant computing resources and bandwidth, and raise privacy concerns due to the continuous transmission of audio data, even when no commands are being issued, leading to inefficient keyword detection and processing.

Innovation Solution

Implementing a two-stage keyword spotting system with a Support Vector Machine (SVM) classifier that reduces the number of feature dimensions and merges support vectors, allowing for efficient keyword detection on local devices, thereby minimizing resource usage and improving processing efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If continuous audio transmission is used for speech recognition, then recognition capability is improved, but bandwidth consumption and computing resources increase significantly

Engineering Contradiction:

Improvespeech recognition capabilityVSAvoidbandwidth consumption

Core Design Contradiction:

ReliabilityVSLoss of energy

Solution Approach 1:

The system performs preliminary keyword spotting locally on the device before transmitting audio data. The wake word detector continuously monitors audio locally for wake words, so that transmission only occurs when a wake word is detected, eliminating continuous transmission and reducing bandwidth consumption while maintaining recognition capability

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The speech recognition system is segmented into two independent stages: local keyword spotting (wake word detection) and remote speech recognition. This segmentation allows the system to perform minimal local processing and only transmit when necessary, reducing overall bandwidth and computing resource usage while preserving full recognition capability when needed

Inventive Principle:

Principle #1Segmentation

2Reliability

If continuous audio transmission is used for speech recognition, then recognition capability is improved, but privacy concerns increase

Engineering Contradiction:

Improvespeech recognition capabilityVSAvoidprivacy concerns

Core Design Contradiction:

ReliabilityVSObject-affected harmful factors

Solution Approach 1:

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

Inventive Principle:

Principle #1Segmentation

3Measurement precision

If full speech recognition processing is performed continuously, then recognition accuracy is improved, but computing resources and processing time increase

Engineering Contradiction:

Improvekeyword detection accuracyVSAvoidprocessing efficiency

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

Inventive Principle:

Principle #1Segmentation

4Productivity

If SVM model size is reduced by reducing support vectors, then processing efficiency is improved, but detection accuracy may deteriorate

Engineering Contradiction:

Improveprocessing efficiencyVSAvoidkeyword detection accuracy

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The patent extracts and removes redundant support vectors from the SVM model while preserving the essential ones needed for accurate classification. By identifying and eliminating unnecessary support vectors, the model size is reduced and processing efficiency is improved without significantly impacting detection accuracy

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter of support vector quantity in the SVM model, reducing it from a large number to a smaller optimized set. This parameter change optimizes the balance between model size and accuracy, improving processing efficiency while maintaining sufficient detection performance

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS9600231B1Model shrinking for embedded keyword spotting

Publication Date: 2017.03.21 AMAZON TECH INC

AI summary

A revised support vector machine (SVM) classifier is offered to distinguish between true keywords and false positives based on output from a keyword spotting component of a speech recognition system. The SVM operates on a reduced set of feature dimensions, where the feature dimensions are selected based on their ability to distinguish between true keywords and false positives. Further, support vectors pairs are merged to create a reduced set of re-weighted support vectors. These techniques result in an SVM that may be operated using reduced computing resources, thus improving system performance.