Open-vocabulary Keyword Spotting with Hyper-Matched Filters for Small Footprint Devices.

Area of Science:

• Artificial Intelligence
• Speech Processing
• Machine Learning

Background:

• Open-vocabulary keyword spotting (KWS) enables the detection of any word in speech, irrespective of the training data.
• Existing KWS models often struggle with efficiency and generalization to diverse speech conditions.
• Small-footprint devices require highly optimized models for on-device keyword detection.

Purpose of the Study:

• To introduce a novel open-vocabulary keyword spotting model designed for high accuracy on small-footprint devices.
• To demonstrate state-of-the-art detection performance with a computationally efficient architecture.
• To evaluate the model's generalization capabilities across different domains and accents.

Main Methods:

• The model utilizes a speech encoder (tiny Whisper or tiny Conformer) and a target keyword encoder implemented as a hyper-network.
• A keyword-specific matched filter is generated by the hyper-network for each target keyword.
• A detection network employs keyword-specific convolution and a Perceiver module with cross-attention for accurate spotting.

Main Results:

• The proposed system achieves state-of-the-art detection accuracy in open-vocabulary keyword spotting.
• The model demonstrates strong generalization to out-of-domain conditions, including second-language (L2) speech.
• The smallest model variant (4.2 million parameters) matches or surpasses larger models in performance.

Conclusions:

• The developed open-vocabulary KWS model offers a compelling balance of efficiency and high accuracy.
• This approach is suitable for deployment on resource-constrained devices requiring robust keyword detection.
• The model's ability to handle unseen words and diverse speech conditions marks a significant advancement.