Wakeword Detection Using Embedding Space Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech-processing systems face challenges in efficiently detecting wakewords due to high computational costs and power consumption, limited flexibility in detecting new wakewords, and potential user experience issues from false wake-ups, especially when predetermined wakewords overlap with user names or common phrases.
Innovation Solution
A user device employs a neural network model trained in a distributed environment to map audio data into an N-dimensional embedding space, allowing for the detection of wakewords and enabling the creation of custom wakewords by processing audio data to determine similarity with predefined wakewords, thereby reducing unnecessary processing and improving user control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speech-processing techniques are used to detect wakewords, then the system can recognize predetermined wakewords, but the computational cost and power consumption are high
Solution Approach 1:
The patent segments the wakeword detection task into two distinct components: (1) a neural network component that maps audio data to embedding space representations, and (2) a separate similarity comparison component that compares embeddings against stored wakeword embeddings. This segmentation allows the computationally intensive neural network to be executed less frequently or with optimized resources, while the lighter similarity comparison handles real-time detection, thereby reducing overall power consumption while maintaining detection accuracy.
Solution Approach 2:
The patent introduces embedding space representations as an intermediary between raw audio data and wakeword detection decisions. Audio data is first transformed into compact embedding vectors that capture essential acoustic features, and then these embeddings are compared against stored wakeword embeddings. This intermediary representation reduces the computational burden of direct audio processing while preserving detection accuracy, thus lowering power consumption.
2Adaptability or versatility
If predetermined wakewords are used for detection, then the system can operate with simple logic, but flexibility to detect new wakewords is limited
Solution Approach 1:
The patent performs preliminary action by pre-computing and storing embedding representations of wakewords in the embedding space during system initialization or training phases. When a new wakeword needs to be detected, the system does not need to reprocess raw audio through the entire neural network; instead, it can directly compare audio embeddings against the pre-computed wakeword embeddings. This preliminary preparation enables flexible detection of new wakewords while keeping runtime processing simple and efficient.
Solution Approach 2:
The patent enables adaptability by allowing the embedding space parameters (wakeword embeddings) to be updated and modified. Instead of being fixed to predetermined wakewords, the system can learn and store embeddings for new wakewords by processing their audio representations through the neural network and saving the resulting embeddings. This parameter flexibility allows the system to adapt to new wakewords without fundamental architectural changes.
3Reliability
If continuous audio processing is performed to ensure accurate wakeword detection, then detection reliability is improved, but false wake-ups increase when wakewords overlap with user names or common phrases
Solution Approach 1:
The patent applies local quality by creating distinct, localized regions in the embedding space for different wakewords. Each wakeword occupies a specific region defined by its embedding representation, and the similarity comparison mechanism determines which region an audio input falls into. This localized embedding space organization improves detection reliability by clearly distinguishing between different wakewords and reduces false wake-ups by preventing overlap between user names/common phrases and actual wakeword regions, as the embedding space naturally separates semantically distinct concepts.
Data Source
AI summary
A neural network model of a user device is trained to map different words represented in audio data to different points in an N-dimensional embedding space. When the user device determines that a mapped point corresponds to a wakeword, it causes further audio processing, such as automatic speech recognition or natural-language understanding, to be performed on the audio data. The user device may first create the wakeword by first processing audio data representing the wakeword to determine the mapped point in the embedding space.