Parametric Pooling Layer for Noise-Robust Wakeword Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing wakeword detection systems for Virtual Personal Assistants (VPAs) face challenges in accurately detecting wakewords in various noise conditions without requiring extensive data collection for each potential sound environment, and they struggle to operate efficiently on resource-constrained devices with minimal latency.
Innovation Solution
The use of deep neural networks with parametric pooling layers that learn to distinguish between informative and non-informative feature vectors, allowing for the aggregation of feature vectors into a compact embedding vector that is robust to noise and can be processed efficiently on resource-constrained devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional pooling layers (average pooling or max pooling) are used to aggregate feature vectors, then the computational complexity is low and the system can operate on resource-constrained devices, but the detection accuracy deteriorates in noisy environments because informative features are diluted or non-informative features are over-emphasized
Solution Approach 1:
The patent transforms the fixed, non-parametric pooling operation into a parametric, learnable operation. The global parametric pooling layer introduces trainable parameters that allow the system to dynamically adjust weighting of feature vectors based on their informativeness, thereby improving detection accuracy without excessive complexity increase.
Solution Approach 2:
The global parametric pooling layer automatically learns to distinguish informative from non-informative feature vectors through training, without requiring manual intervention or extensive external data. The system self-optimizes the pooling weights based on the features it receives, enabling adaptability to different acoustic environments.
2Reliability
If extensive data is collected for each potential sound condition to improve detection robustness, then the wakeword detection becomes more robust to noise, but the data collection time and system complexity increase substantially
Solution Approach 1:
The global parametric pooling layer serves as a universal component that can handle various acoustic conditions without requiring separate models or extensive condition-specific data. It learns general patterns of informative versus non-informative features that apply across different noise environments, enabling the system to generalize to unseen conditions.
Solution Approach 2:
The system performs preliminary learning of feature weighting patterns during model training, so that when deployed, it can immediately apply these learned patterns to new acoustic conditions without requiring additional data collection or adaptation time in the field.
3Speed
If the VPA continuously monitors acoustic signals for wakeword detection, then the response time to user commands is minimized, but the energy consumption and computational load on resource-constrained devices increase
Solution Approach 1:
The global parametric pooling layer efficiently aggregates variable-length feature sequences into fixed-length embeddings, enabling continuous monitoring with optimized computational parameters. This reduces the computational burden per time step while maintaining the ability to detect wakewords in real-time.
Solution Approach 2:
The system maintains continuous wakeword detection capability through the efficient global parametric pooling operation, which processes incoming acoustic features without interruption. The learned parametric weights enable continuous operation with reduced computational overhead compared to traditional pooling methods.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The current disclosure relates to systems and methods for wakeword or keyword detection in Virtual Personal Assistants (VPAs). In particular, systems and methods are provided for wakeword detection using deep neural networks including a parametric pooling layer, wherein the parametric pooling layer includes trainable parameters, enabling the layer to learn to distinguish between informative feature vectors and non-informative/noisy feature vectors extracted from a variable length acoustic signal. In one example, a parametric pooling layer may aggregate a variable length feature map, comprising a plurality of feature vectors extracted from an acoustic signal, into an embedding vector of pre-determined length, by weighting each of the plurality of feature vectors based on one or more learned parameters in a parametric pooling layer, and aggregating the plurality of weighted feature vectors into the embedding vector.