Optimizing Keyword Spotting Classifier Based on Tiny Machine Learning for Low-Power Embedded Devices

Keyword Spotting (KWS) is essential for enabling voice recognition in smart home systems and although deploying reliable and energy-efficient models on low-power embedded devices remains a challenge, tiny machine learning offers a promising solution. In this paper, raw audio data from Google's Speech Commands Dataset <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{v} 0.02$</tex> is used to develop and evaluate neural network architectures optimized for KWS based on Mel-frequency cepstral coefficients (MFCC). The audio signals are formatted as a compact 2 D matrix of time-frequency features, making it ideal for convolutional neural networks (CNNs). Ten networks are trained with 10-fold cross-validation, fine-tuned on speaker-specific data and, by using post-training quantization, converted from TensorFlow (TF) to TensorFlow Lite (LiteRT) format for deployment on microcontrollers. The trade-offs between model size, performance and energy efficiency are analyzed, and the results show that LiteRT models considerably reduce the energy consumption, with energy savings ranging from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1-4 \%$</tex> for smaller capacity models to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 9. 7 3 \%}$</tex> for larger ones, while maintaining performance similar to TF models. Furthermore, fine-tuning models of greater capacity slightly improves their accuracy (0.51%) and energy efficiency (1.75%), whereas smaller models are unaffected by fine-tuning. This study lays the foundation for the integration of KWS models based on tiny machine learning into microcontrollers for real-time applications.