Dynamic Wakeword Spotter for Speech Devices

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech-enabled devices face challenges in customizing factory-installed wakewords, leading to collisions and poor user experiences due to the inability to quickly train new wakeword spotters without large datasets of audio samples.

Innovation Solution

A method is provided to modify wakewords by receiving a spoken request, parsing it into a natural language request and a speech audio segment, and building a new wakeword spotter to recognize the new wakeword as an activation trigger, using approaches such as wakeword segmentation, sampling, or continuous transcription.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If factory-installed wakewords are used, then devices can be deployed immediately, but multiple devices may collide and activate simultaneously leading to poor user experience

Engineering Contradiction:
Improvedevice deployment speedVSAvoidactivation accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system changes the wakeword parameter from a fixed factory-default value to a dynamically customizable value. Users can modify the wakeword to be unique across multiple devices, eliminating collisions while maintaining immediate device functionality. This parameter change allows the same device model to have different wakewords, resolving the contradiction between rapid deployment and reliable activation.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If traditional wakeword spotter training is used, then accurate wakeword recognition is achieved, but large datasets of audio samples are required which cannot be obtained for dynamic wakewords

Engineering Contradiction:
Improvewakeword recognition accuracyVSAvoidaudio sample data volume
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The system performs preliminary action by pre-training wakeword spotters using factory-default wakewords and their associated audio datasets before device deployment. When users customize wakewords, the system leverages these pre-trained models as a foundation, requiring only minimal adaptation rather than complete retraining. This preliminary preparation enables accurate recognition of dynamic wakewords without needing large custom datasets for each user configuration.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates a copy of the pre-trained wakeword spotter model and adapts it for the user's custom wakeword. Instead of training from scratch, the system copies the existing model's knowledge of wakeword patterns and acoustic features, then fine-tunes it for the specific custom wakeword. This copying approach maintains high recognition accuracy while minimizing the need for extensive new training data.

Inventive Principle:
Principle #26Copying

3Adaptability or versatility

If wakeword spotters are trained immediately without large datasets, then dynamic wakewords can be used, but traditional training approaches are not available

Engineering Contradiction:
Improvewakeword customization capabilityVSAvoidspotter training feasibility
Core Design Contradiction:
Adaptability or versatilityVSEase of manufacture

Solution Approach 1:

The system implements self-service by automatically generating training data and adapting models without requiring manual data collection or extensive user input. When a user customizes a wakeword, the system automatically uses the user's voice recording during the customization process as training data, and automatically adapts the spotter model. This self-service mechanism makes immediate spotter training feasible for dynamic wakewords by eliminating the need for separate data collection and manual training procedures.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4550313A1Dynamic wakewords for speech-enabled devices
Publication Date: 2025.05.07 SOUNDHOUND INC
  • EP4550313A1 patent drawingFigure 1
  • EP4550313A1 patent drawingFigure 2
  • EP4550313A1 patent drawingFigure 3

AI summary

The present disclosure relates to dynamic wakewords for speech-enabled devices. A system and method are disclosed capable of parsing a spoken utterance into a natural language request and a speech audio segment, where the natural language request directs the system to use the speech audio segment as a new wakeword. In response to this wakeword assignment directive, the system and method are further capable of immediately building a new wakeword spotter to activate the device upon matching the new wakeword in the input audio. Different approaches to promptly building a new wakeword spotter are described. Variations of wakeword assignment directives can make the new wakeword public or private. They can also add the new wakeword to earlier wakewords, or replace earlier wakewords.