Shared-Encoder Wakeword Detection for Multi-Keyword Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition systems face challenges in efficiently detecting multiple wakewords due to the high computational resource requirements of adding new wakeword detectors, particularly in devices with limited resources, making it impractical to operate multiple wakeword detectors simultaneously.

Innovation Solution

The system employs a two-layer architecture with a joint encoder for feature extraction and multiple decoders for wakeword detection, allowing devices to recognize multiple wakewords without overwhelming computing resources by training new decoders for each wakeword, reusing the joint encoder.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If multiple traditional wakeword detectors are deployed simultaneously, then multiple wakewords can be detected, but computational resource requirements increase significantly

Engineering Contradiction:
Improvewakeword detection capabilityVSAvoidcomputational resource consumption
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent merges multiple wakeword detection models into a single unified neural network model. The encoder shares common feature extraction capabilities across all wakewords, while only the decoder portions differ for each wakeword. This combining approach reduces the total number of parameters and computational operations compared to running separate detectors, directly addressing the resource consumption problem while maintaining multi-wakeword detection capability.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The encoder component serves as a universal feature extractor that processes audio input for all wakewords simultaneously. By designing the encoder to be shared across multiple wakeword detection tasks, the system achieves multi-functionality where a single component performs the same feature extraction role for different wakewords, reducing redundancy and computational overhead.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Adaptability or versatility

If multiple separate wakeword detectors are operated, then different wakewords can be recognized, but device complexity increases

Engineering Contradiction:
Improvewakeword recognition capabilityVSAvoiddetector architecture complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent combines multiple detector architectures into a single integrated model structure. By merging the encoding stage into a shared component and only duplicating the decoding stage for each wakeword, the system reduces architectural complexity while preserving the ability to recognize multiple wakewords. This unified structure is simpler to deploy and manage than multiple independent detectors.

Inventive Principle:
Principle #5Merging (Combining)

3Adaptability or versatility

If traditional wakeword detection systems are updated with new wakewords, then wakeword versatility improves, but training computational requirements increase

Engineering Contradiction:
Improvewakeword varietyVSAvoidtraining computational power
Core Design Contradiction:
Adaptability or versatilityVSPower

Solution Approach 1:

The patent segments the wakeword detection model into two distinct parts: a shared encoder and wakeword-specific decoders. This segmentation allows independent training of components - the encoder can be trained once on general audio features, and new wakewords can be added by training only their specific decoder portions. This dramatically reduces training computational requirements when adding new wakewords compared to retraining entire separate detectors.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The encoder is trained in advance on general audio feature extraction tasks that are common to all wakewords. This preliminary training of the shared encoder means that when new wakewords need to be added, the system only needs to train the new decoder portions rather than retraining the entire detection pipeline, significantly reducing the computational power required for updates.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4445364B1Multiple wakeword detection
Publication Date: 2026.03.18 AMAZON TECH INC
  • EP4445364B1 patent drawingFigure 1A
  • EP4445364B1 patent drawingFigure 1B
  • EP4445364B1 patent drawingFigure 2

AI summary

A device is configured to detect multiple different wakewords. A device may operate a joint encoder that operates on audio data to determine encoded audio data. The device may operate multiple different decoders which process the encoded audio data to determine if a wakeword is detected. Each decoder may correspond to a different wakeword. The decoders may use fewer computing resources than the joint encoder, allowing for the device to more easily perform multiple wakeword processing. Enabling / disabling wakeword(s) may involve the reconfiguring of a wakeword detector to add / remove data for respective decoder(s).