Speech Keyword Spotting on CIM Chips for Low-Power Edge Control

Solution Overview

Problem

Intelligent devices face significant resource consumption issues, particularly in power supply and storage space, due to real-time speech monitoring for keyword spotting, which affects service time and efficiency.

Innovation Solution

A speech data processing method utilizing a programmable logic unit for feature extraction, converting to a multi-channel input feature map, and performing keyword spotting on a computer-in-memory chip using a ConvMixer architecture-based model, integrated with batch normalization and reduced parameter layers, to optimize power and resource usage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If real-time speech monitoring is performed for keyword spotting, then speech recognition capability is improved, but power consumption and storage space increase significantly

Engineering Contradiction:

Improvespeech recognition capabilityVSAvoidpower consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The speech processing system is segmented into two distinct parts: a lightweight keyword spotting module that runs continuously on the edge device with minimal resource consumption, and a full speech recognition model that only activates when keywords are detected. This segmentation allows the system to maintain speech recognition capability while dramatically reducing power consumption during idle periods.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The keyword spotting model performs preliminary filtering of speech data before triggering the more resource-intensive speech recognition model. By pre-processing the audio stream with a lightweight model that continuously monitors for wake words or command keywords, the system prepares and identifies relevant speech events without committing full computational resources until necessary.

Inventive Principle:

Principle #10Preliminary action

2Measurement precision

If real-time speech monitoring is performed for keyword spotting, then speech recognition capability is improved, but storage space consumption increases

Engineering Contradiction:

Improvespeech recognition capabilityVSAvoidstorage space

Core Design Contradiction:

Measurement precisionVSQuantity of substance

Solution Approach 1:

The speech processing system is segmented into two distinct parts: a lightweight keyword spotting module that runs continuously on the edge device with minimal resource consumption, and a full speech recognition model that only activates when keywords are detected. This segmentation allows the system to maintain speech recognition capability while dramatically reducing power consumption during idle periods.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The keyword spotting model performs preliminary filtering of speech data before triggering the more resource-intensive speech recognition model. By pre-processing the audio stream with a lightweight model that continuously monitors for wake words or command keywords, the system prepares and identifies relevant speech events without committing full computational resources until necessary.

Inventive Principle:

Principle #10Preliminary action

3Measurement precision

If traditional speech recognition models are used, then recognition accuracy is improved, but processing delay and power consumption increase

Engineering Contradiction:

Improverecognition accuracyVSAvoidprocessing delay

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The speech processing system is segmented into two distinct parts: a lightweight keyword spotting module that runs continuously on the edge device with minimal resource consumption, and a full speech recognition model that only activates when keywords are detected. This segmentation allows the system to maintain speech recognition capability while dramatically reducing power consumption during idle periods.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system employs periodic action by having the lightweight keyword spotting model continuously monitor audio input at low computational cost, then periodically activating the heavier speech recognition model only when keywords are detected. This periodic activation pattern reduces average processing delay while maintaining recognition accuracy for relevant speech events.

Inventive Principle:

Principle #19Periodic action

Data Source

PatentUS20260031084A1Speech data processing method, speech data processing device and speech control system

Publication Date: 2026.01.29 BEIJING BOE TECH DEV CO LTD

AI summary

The present disclosure relates to the technical field of computers, and provides a speech data processing method and device. The method includes: acquiring speech data to be processed; sending said speech data to a programmable logic unit to perform speech feature extraction on said speech data, to acquire a speech feature vector corresponding to said speech data; in programmable logic unit, converting the speech feature vector into a multi-channel input feature map; sending the input feature map to a computer-in-memory chip, and performing speech keyword spotting on the input feature map by using a pre-configured speech keyword spotting model on the computer-in-memory chip to obtain a speech keyword spotting result; and feeding back the speech keyword spotting result to a main control system, the main control system is configured to execute, according to the speech keyword spotting result, a response operation corresponding to the speech keyword spotting result.