[Paper Review] Low-Power Audio Keyword Spotting using Tsetlin Machines

Jie Lei, Tousif Rahman+6 more|arXiv (Cornell University)|Jan 27, 2021

Music and Audio Processing43 references4 citations

TL;DR

This paper proposes a low-power keyword spotting (KWS) system using Tsetlin Machines (TMs), a logic-based machine learning model that replaces arithmetic-heavy operations in neural networks with Boolean logic, enabling faster convergence, fewer parameters, and lower energy consumption. The TM-based KWS pipeline achieves comparable accuracy to DNNs with significantly reduced complexity and hardware footprint, demonstrating strong potential for on-chip deployment in edge AI devices.

ABSTRACT

The emergence of Artificial Intelligence (AI) driven Keyword Spotting (KWS) technologies has revolutionized human to machine interaction. Yet, the challenge of end-to-end energy efficiency, memory footprint and system complexity of current Neural Network (NN) powered AI-KWS pipelines has remained ever present. This paper evaluates KWS utilizing a learning automata powered machine learning algorithm called the Tsetlin Machine (TM). Through significant reduction in parameter requirements and choosing logic over arithmetic based processing, the TM offers new opportunities for low-power KWS while maintaining high learning efficacy. In this paper we explore a TM based keyword spotting (KWS) pipeline to demonstrate low complexity with faster rate of convergence compared to NNs. Further, we investigate the scalability with increasing keywords and explore the potential for enabling low-power on-chip KWS.

Motivation & Objective

• To address the high energy and computational complexity of neural network-based keyword spotting (KWS) in edge devices.
• To explore the feasibility of using Tsetlin Machines (TMs) as a low-power alternative to deep neural networks (DNNs) in KWS pipelines.
• To evaluate the scalability and energy efficiency of the TM-based KWS system under varying hyperparameters and keyword counts.
• To demonstrate the potential for hardware acceleration of the KWS pipeline using logic-based processing instead of arithmetic computation.
• To enable real-time, on-chip KWS by minimizing model size and system latency through TM-specific optimizations.

Proposed method

• The KWS pipeline uses Mel-frequency cepstrum coefficients (MFCCs) as input features, which are discretized into Boolean form using quantile binning.
• The Tsetlin Machine processes these Boolean features using a clause-based logic architecture, replacing gradient-based backpropagation with feedback-driven learning automata.
• Hyperparameters such as the number of clauses, clause size, and threshold values are tuned to balance accuracy, energy efficiency, and convergence speed.
• The system is implemented on a Raspberry Pi to evaluate real-time performance and energy consumption in a low-power embedded environment.
• The pipeline is compared against standard DNN and quantized CNN models using the TensorFlow Speech Commands dataset.
• Signal processing optimizations include window function adjustments to reduce MFCC dimensionality with minimal performance loss.

Experimental results

Research questions

• RQ1 Can a Tsetlin Machine-based KWS pipeline achieve comparable accuracy to deep neural networks while using significantly fewer parameters?
• RQ2 How does the number of clauses and feedback hyperparameters affect the convergence speed and energy efficiency of the TM-based KWS system?
• RQ3 To what extent can MFCC discretization be simplified (e.g., to one Boolean per feature) without degrading classification performance?
• RQ4 How does the TM-based KWS pipeline scale with increasing numbers of keywords in terms of accuracy and resource usage?
• RQ5 Can the TM-based system be effectively deployed on low-power microcontrollers to enable on-chip, real-time keyword spotting?

Key findings

• The Tsetlin Machine-based KWS pipeline achieved comparable accuracy to DNNs and quantized CNNs while using orders of magnitude fewer parameters and clauses.
• The system demonstrated faster convergence during training compared to traditional neural networks, due to the absence of gradient descent and arithmetic-intensive operations.
• Reducing MFCC features to one Boolean per coefficient had minimal impact on performance, enabling significant reduction in model size and computational load.
• Lowering the number of clauses and tuning the threshold hyperparameter improved energy efficiency and latency without sacrificing accuracy.
• The pipeline showed strong potential for hardware acceleration due to its logic-based computation, which is inherently more energy-efficient than floating-point arithmetic.
• The system was successfully evaluated on a Raspberry Pi, confirming its viability for low-power, on-chip deployment in edge AI applications.