Dozens, Not Thousands: A Small Dataset Pipeline for Custom Wake Words

Abstract

We introduce an end-to-end pipeline for customizable wake word detection which does not rely on pre-existing dataset and is designed for robotics and smart interactive devices. By leveraging the pre-trained OpenAI Whisper encoder and a LoRA adapter, our method can be trained with only a few dozen user-provided audio samples, removing the need for large datasets typically required by state-of-the-art systems. The automated pipeline handles sample collection, synthetic audio generation, training, and real-time evaluation, achieving low false acceptance and rejection rates while maintaining efficient GPU inference. We validate the framework on three different wake words and evaluate its performance through real-time evaluation runs. Our results highlight the system's potential for deployment of personalized, high-accuracy voice activation in robotics and other branded or individualized agent solutions, while using a fraction of the data needed by other frameworks.