The Tsetlin Machine (TM) has recently attracted attention as a low-power alternative to neural networks due to its simple and interpretable inference mechanisms. However, its performance on speech-related tasks remains limited. This paper proposes TsetlinKWS, the first algorithm-hardware co-design framework for the Convolutional Tsetlin Machine (CTM) on the 12-keyword spotting task. Firstly, we introduce a novel Mel-Frequency Spectral Coefficient and Spectral Flux (MFSC-SF) feature extraction scheme together with spectral convolution, achieving a record-high accuracy of 87.35% for TM-based models on the 12-keyword spotting task. Secondly, we develop an Optimized Grouped Block-Compressed Sparse Row (OG-BCSR) algorithm that delivers a remarkable 9.84× reduction in model size, significantly improving the storage efficiency of CTMs. Finally, we propose a state-driven architecture tailored for CTMs, which simultaneously exploits data reuse and sparsity to attain high energy efficiency. The full system is evaluated in 65 nm process technology, consuming 16.58 μW at 0.7 V with a compact 0.63 mm2 core area. TsetlinKWS demonstrates a Power-Performance-Area (PPA) product of 60.2 and requires only 907k logic operations per inference, corresponding to 1.4× and 10× reductions, respectively, compared with state-of-the-art KWS accelerators. These results position the CTM as a highly efficient candidate for ultra-low-power speech applications.
Figures & tables
Fig. 1: NN-based KWS accelerator categories. (a) Audio signal. (b) Batch-based accelerators are represented by convolutional neural networks [ 1 ] , [ 9 ] , [ 10 ] , [ 11 ] , [ 12 ] , [ 13 ] . They perform inference in a multi-frame manner because neurons do not have states. (c) Frame-based accelerators are represented by recurrent neural networks [ 5 ] , [ 14 ] , [ 15 ] , [ 16 ] . They process only one frame of features per inference.
Fig. 2: The structure of the Tsetlin Machine. Xk refers to the Boolean input. Included TAs and excluded TAs determine the matching pattern of clause Cij .
Fig. 3: Two common energy-efficient edge accelerator architectures. (a) Data-driven architecture: efficient for data reuse but inflexible for sparse acceleration. (b) Event-driven architecture: efficient for sparse acceleration but limited in data reuse.
Fig. 4: The system architecture of TsetlinKWS.
Fig. 5: Feature extraction process: (a) Traditional MFCC flow. (b) The proposed MFSC-SF flow. An additional overlap difference is proposed to achieve the calculation of spectral flux, and some sub-modules are deleted.
Fig. 6: The architecture diagram of the MFSC-SF feature extractor. MFSC and SF are obtained in the overlap unit and then are binarized in the binarizer. MFSC is binarized using the dynamic mean, while SF is binarized using a fixed threshold.
Fig. 7: Accuracy of CTMs under different pre-processing algorithms and at different model scales. “32/64” represents the number of filter bank channels. The 120-Clause CTM with MFSC32-SF32 is the selected model in this work.
Fig. 8: Example of compressing the TA action matrix using OG-BCSR. The block index is used to skip storage of empty blocks, e.g., the first block B0 is empty. The row count indicates the number of included TAs in each row. The column index indicates their column addresses.
Fig. 9: Construction process from the TA action matrix to the graph. The weight of edges is the number of empty blocks after merging. In the example, n0-n1 and n2-n3 form the optimal matching.
Fig. 10: Compression results of CSR and OG-BCSR algorithms.
Fig. 11: The spectral convolution kernel covers the entire spectral coefficient dimension to achieve efficient feature matching. Two smiling faces denote the advantages of the spectral convolution kernel in terms of performance and power, while a neutral face represents an acceptable area trade-off. Wi refers to the weight of the convolution kernel Ki .
Fig. 12: The computation process of the CTM and the state-driven strategy. The sliding literal data is obtained according to the position of included TAs, and it is directly computed to equivalently achieve the convolution result.
Fig. 13: The core pipeline architecture in TsetlinKWS. The OG-BCSR decompression circuit comprises three stages (S1–S3), followed by the PE array (S4). The ‘×5’ notation denotes five parallel priority encoding units.
Fig. 14: Example of a two-stage scheduling optimization strategy.
Fig. 15: Comparisons of the execution cycle and PE utilization in each optimization stage.
Fig. 16: Comparison of normalized memory-access costs (decomposed into input activations, weights, and output activations) and inference latencies across different architectures or TA sparsity levels.
Fig. 17: (a) The layout of TsetlinKWS (exported from Cadence Virtuoso). (b) The performance metrics of TsetlinKWS.
Fig. 18: Power breakdown of TsetlinKWS, feature extractor, and CTM accelerator at 0.7 V (measured on Synopsys PTPX).
Batch-based Accelerator
Frame-based Accelerator
This work
OJSSCS’23 [ 9 ]
TCAS-I’26 [ 13 ]
JSSC’20 [ 15 ]
TCASAI’24 [ 16 ]
Technology
65 nm
22 nm
28 nm
65 nm
65 nm
Voltage (V)
0.7
0.54
0.42
0.6
0.6/0.65
Core Area (mm 2 )
0.63
(FEx: 50%, Accel: 50%)
0.6
(FEx: 42%, Accel: 58%)
0.37
TABLE I: Comparison of SoTA KWS Works
Kernel Size
#Clause/Class
Training Accuracy @20-40dB
Test Accuracy @clean
64×3
240
85.73%
86.55%
64×7
120
85.33%
87.35%
64×15
60
82.83%
84.89%
TABLE II: Experimental Results for Different Kernel Sizes
TABLE III: Grid Search Results for Hyperparameters
Keyword Spotting (KWS) is becoming increasingly important as voice-controlled devices grow more widespread. While voice interaction with smartphones and smart TVs is already common, deploying KWS on heavily resource-constrained edge devices such as wearables remains challenging. These systems must meet high accuracy requirements while operating under strict constraints on computational power, memory footprint, and real-time latency. In this work, we present an application of the STMC (Short-Term Memory Convolutions) framework to adapt a modular CNN model for online, LSTM-like inference. Our approach reduces power consumption and redundant computations while maintaining the stability and simplicity of training CNNs. We achieve up to 82% and 46% MCPS reduction compared to equivalently frequent standard CNN execution and vanilla STMC, respectively. The best configuration achieves 93.8% accuracy on the 11-class Google Speech Commands task and 97.1% on the same task with zero-padded data.
Paweł Warlewski, Artur Czeczko, Artur Szumaczuk +2
Samsung R&D Institute Poland · Samsung AI Center Warsaw
Keyword spotting (KWS), the task of identifying predefined words in speech, is a core capability of voice-enabled devices. Achieving high KWS accuracy under tight parameter budgets across different vocabulary sizes remains challenging. We present CircleMatch, a matching framework enabling KWS with very few parameters. Its encoder independently compresses frequency bands and fuses them into frame features. These features are then matched against learned class-specific prototypes to produce temporal response curves. Parameter-free circular aggregation encodes time as angles and summarizes response distributions and relative timing for classification. We develop four tiny variants, Circle-D4, Circle-D8, Circle-D16, and Circle-D32, ranging from approximately 1k to 7k parameters in the 12-class setting. Experiments with multiple random seeds on Speech Commands v1/v2 and the English and Spanish Micro subsets of the Multilingual Spoken Words Corpus demonstrate competitive accuracy with tiny models. Our qualitative analysis further suggests approximate shift equivariance of prototype responses and adaptation to temporal compression. Code and model weights are available at https://github.com/ora942878/CircleMatch.
A very low complexity feature extractor called next iRDT is proposed and evaluated for the problem of keyword spotting (KWS). Unlike any other types of feature extractors including the widely used MFCC, or adaptive, CNN-based ones, our algorithm is multiplier-free and it employs only simple, energy-efficient arithmetic operators. Since keyword-spotting of speech commands (KWS) is a typical application for TinyML platforms requiring low complexity for the signal classification chain, we consider it as a case study to evaluate complexity and functional performance. If properly tuned, iRDT demonstrates similar accuracy to solutions based on MFCC or CNN-based extractors using baseline classifiers on Google's KWS 12-classes dataset. With a different classifier the system achieved 94.7% validation accuracy. Processing times on CPU for the proposed feature extractor, are at least one order of magnitude smaller than for the MFCC. The proposed algorithm has a very low hardware footprint, making it ideal for ultra-low power edge devices. Code and demo are available [18].
Radu Dogaru, Ioana Dogaru
Dept. of Applied Electronics and Information Engineering, National University of Science and Technology POLITEHNICA Bucharest, Romania