A Simplified Pipeline for Wakeword Creation and Deployment: Leveraging Zero-Shot Text-to-Speech and ROS2 for Robotic Systems

Wake word detection is a critical component for voice-activated robotic systems. However, the lack of diverse datasets and Text-to-Speech (TTS) models in non-English languages, such as Portuguese, poses significant challenges for developing accurate models. This paper proposes a novel and streamlined pipeline for the creation of wake word models, inspired by the OpenWakeWord framework, which leverages synthetically generated datasets to train neural models. Unlike existing approaches, our method employs a zero-shot TTS model capable of producing high-quality audio across 17 languages and a broad range of speaker styles, effectively simulating a multi-speaker environment. In contrast to OpenWakeWord, which relies on English-centric data and fixed voice banks, our approach introduces multilingual support and greater speaker variability without requiring speaker enrollment. We also detail the integration of this pipeline with the Robot Operating System 2 (ROS 2), enabling real-time robotic applications. Experimental results show that our models achieve an F1-score of up to 0.91 using purely synthetic data, demonstrating the viability and effectiveness of our method. This work highlights the potential of synthetic data generation to advance voice interaction in languages with limited resources, with a particular focus on the field of robotics in Brazil.