Utterance-Level Wake on Intent Keyword Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing automatic speech recognition (ASR) systems that rely on wake-on-keyphrase technology are unnatural for users and inefficient for platforms with limited compute and memory resources, and they struggle with accurate intent recognition in noisy environments, especially when requiring continuous large vocabulary processing.

Innovation Solution

A spoken keyword detection system that uses utterance-level wake on intent technology, which detects keywords in a sequence to determine intent and triggers ASR only when necessary, employing neural networks and feature representations like MFCC, phones, WORD2VEC, and SPEECH2VEC for intent classification, suitable for low-power devices.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If wake-on-keyphrase technology is used, then the system can detect specific keyphrases, but it requires continuous large vocabulary processing which consumes excessive power and memory resources

Engineering Contradiction:

Improvekeyword detection accuracyVSAvoidpower consumption

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent segments the speech processing task into two stages: first, a lightweight keyword spotting model detects potential keywords with low computational cost; second, a more accurate ASR model processes only the segments containing detected keywords. This segmentation allows the system to maintain keyword detection accuracy while dramatically reducing power consumption by avoiding continuous full ASR processing.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system performs partial action by only triggering the computationally intensive ASR model when keywords are detected by the lightweight model. Instead of continuously processing the entire vocabulary, the system applies full processing power selectively and partially, only when necessary, thereby reducing overall power consumption while maintaining detection accuracy.

Inventive Principle:

Principle #16Partial or excessive action

2Measurement precision

If wake-on-keyphrase technology is used, then the system can identify specific keyphrases, but it struggles with accurate intent recognition in noisy environments

Engineering Contradiction:

Improveintent recognition accuracyVSAvoidnoise interference

Core Design Contradiction:

Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent introduces an intermediary lightweight keyword spotting model that acts as a filter between the noisy environment and the full ASR system. This intermediary model is specifically trained to detect keywords robustly in noisy conditions, and only passes segments containing detected keywords to the full ASR model for intent recognition, thereby protecting the system from noise interference while maintaining accurate intent recognition.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Adaptability or versatility

If continuous large vocabulary processing is performed, then the system can handle diverse speech inputs, but it is inefficient for platforms with limited compute and memory resources

Engineering Contradiction:

Improvespeech processing capabilityVSAvoidprocessing efficiency

Core Design Contradiction:

Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent segments the speech signal into processed segments (containing detected keywords) and unprocessed segments. The lightweight model quickly scans the entire input to identify relevant segments, and only these segments are passed to the full ASR model for detailed processing. This segmentation approach maintains the system's ability to handle diverse speech inputs while dramatically improving processing efficiency by avoiding redundant computation on irrelevant segments.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system applies full vocabulary processing capability partially, only to the portions of speech that contain detected keywords. The lightweight model enables the system to have the adaptability of large vocabulary processing available when needed, but actually exercises this capability only partially and selectively, thereby improving processing efficiency on resource-constrained platforms.

Inventive Principle:

Principle #16Partial or excessive action

4Ease of operation

If fixed keyphrase requirement is enforced, then the system can simplify wake-up detection, but it creates an unnatural user experience

Engineering Contradiction:

Improvewake-up detection simplicityVSAvoiduser interaction flexibility

Core Design Contradiction:

Ease of operationVSAdaptability or versatility

Solution Approach 1:

The patent implements a dynamic two-stage processing system where the lightweight keyword spotting model provides simple, fast wake-up detection, and the full ASR model dynamically processes segments containing detected keywords to enable natural language intent recognition. This dynamic architecture allows the system to maintain simple wake-up detection while gaining the flexibility to understand diverse natural language inputs, creating a more natural user experience.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS10984783B2Spoken keyword detection based utterance-level wake on intent system

Publication Date: 2021.04.20 TAHOE RES LTD

AI summary

An embodiment of a wake-on-intent speech recognition device includes technology to detect one or more keywords in a digital representation of a spoken natural language utterance, determine an intent of the spoken natural language utterance based on the detected keywords, and provide the spoken natural language utterance to a speech recognition and interpretation system if the determined intent is to further process the spoken natural language utterance. Other embodiments are disclosed and claimed.