Custom Wake Word Verification with Universal Acoustic Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing systems face challenges in accurately detecting custom wake words with high false activations and low correct activations, particularly on devices with limited memory and processing bandwidth, and require specialized training for each wake word.

Innovation Solution

A system utilizing a generalized speech recognition model tailored for wake word verification, incorporating a decoding graph and background language model, with neural network quantization and beam search decoder, to efficiently detect user-defined wake words without retraining, reducing computational and power requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a specialized wake word detection model is trained for each custom wake word, then detection accuracy is improved, but device complexity and training requirements increase

Engineering Contradiction:
Improvewake word detection accuracyVSAvoidmodel training complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies a universal wake word detection model that can recognize multiple custom wake words without requiring separate specialized training for each wake word. The model uses a generalized acoustic model combined with a decoding graph that includes the custom wake word and alternate pronunciations, allowing one model to serve multiple wake word detection functions.

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

Solution Approach 2:

The system changes the approach from training separate models for each wake word to using a single universal model with modified parameters. The decoding graph is configured with the specific custom wake word and its alternate pronunciations, allowing the same model architecture to adapt to different wake words by adjusting the decoding graph parameters rather than retraining the entire model.

Inventive Principle:
Principle #35Parameter changes

2Device complexity

If a universal wake word detection model is used for multiple wake words, then device complexity is reduced, but detection accuracy for custom wake words deteriorates

Engineering Contradiction:
Improvemodel training complexityVSAvoidwake word detection accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the wake word detection process into two components: a universal acoustic model that handles audio processing and a decoding graph that is specifically configured for the custom wake word. This segmentation allows the acoustic model to remain universal while the decoding graph provides wake word-specific accuracy through its tailored structure including alternate pronunciations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The decoding graph acts as an intermediary between the universal acoustic model and the wake word detection output. It receives phoneme probabilities from the acoustic model and uses the custom wake word and alternate pronunciations configuration to generate wake word-specific detection results, mediating between general acoustic processing and specific wake word recognition.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Adaptability or versatility

If wake word verification is performed on devices with limited memory and processing bandwidth, then device accessibility is improved, but computational efficiency deteriorates

Engineering Contradiction:
Improvedevice accessibilityVSAvoidcomputational efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent extracts the wake word verification function from the main audio processing pipeline and performs it as a separate verification step. The universal wake word detection model processes audio and generates wake word probabilities, which are then verified against the custom wake word configuration. This extraction allows the system to operate on devices with limited resources by separating the computational tasks.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system applies partial action by using a universal model that processes audio for wake word detection but only performs full verification when the decoded phonemes match the custom wake word configuration. This partial verification approach reduces unnecessary computational overhead while maintaining high accuracy for actual wake word detections.

Inventive Principle:
Principle #16Partial or excessive action

4Speed

If the wake word detection system processes audio in real-time, then response time is improved, but processing bandwidth requirements increase

Engineering Contradiction:
Improveresponse timeVSAvoidprocessing bandwidth
Core Design Contradiction:
SpeedVSPower

Solution Approach 1:

The patent performs preliminary action by pre-configuring the decoding graph with the custom wake word and alternate pronunciations before audio processing. This pre-configuration allows the system to quickly match decoded phonemes against the pre-established wake word patterns during real-time audio processing, reducing the need for complex real-time verification and lowering processing bandwidth requirements.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3966813B1Online verification of custom wake word
Publication Date: 2025.09.10 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3966813B1 patent drawingFigure 1
  • EP3966813B1 patent drawingFigure 2
  • EP3966813B1 patent drawingFigure 3

AI summary

Generally discussed herein are devices, systems, and methods for wake word verification. A method can include receiving, at a server, a message from a device indicating that an utterance of a user-defined wake word was detected at the device, the message including (a) audio samples or features extracted from the audio samples and (b) data indicating the user-defined wake word, retrieving or generating, at the server, a custom decoding graph for the user-defined wake word, wherein the decoding graph and the static portion of the wake word verification model form a custom wake word verification model for the user-defined wake word, executing the wake word verification model to determine a likelihood that the wake word was uttered, and providing a message to the device indicating whether wake was uttered based on the determined likelihood.