Implementing Keyword Spotting on the MCXN947 Microcontroller with Integrated NPU

Petar Jakuš; Hrvoje Džapo

cs.HCJun 10, 2025

Implementing Keyword Spotting on the MCXN947 Microcontroller with Integrated NPU

Authors: Petar Jakuš, Hrvoje Džapo

Organizations: Faculty of Electrical Engineering and Computing, Zagreb, Croatia

Abstract

This paper presents a keyword spotting (KWS) system implemented on the NXP MCXN947 microcontroller with an integrated Neural Processing Unit (NPU), enabling real-time voice interaction on resource-constrained devices. The system combines MFCC feature extraction with a CNN classifier, optimized using Quantization Aware Training to reduce model size with minimal accuracy drop. Experimental results demonstrate a 59x speedup in inference time when leveraging the NPU compared to CPU-only execution, achieving 97.06% accuracy with a model size of 30.58 KB, demonstrating the feasibility of efficient, low-power voice interfaces on embedded platforms.

Figures & tables

Figure 1: MFCC original signal and frames.

Figure 2: Frames after applying Hamming function.

Layer (type)	Output Shape	Param #
input (InputLayer)	(None, 98, 20, 1)	0
conv1 (Conv2D)	(None, 96, 18, 32)	320
bn1 (BatchNormalization)	(None, 96, 18, 32)	128
pool1 (MaxPooling2D)	(None, 48, 9, 32)	0
conv2 (Conv2D)	(None, 46, 7, 64)	18496
bn2 (BatchNormalization)	(None, 46, 7, 64)	256

Table I: Regular Neural Network Architecture

Layer (type)	Output Shape	Param #
input (InputLayer)	(None, 98, 20, 1)	0
quantize_layer (QuantizeLayer)	(None, 98, 20, 1)	3
quant_conv1 (QuantizeWrapperV2)	(None, 96, 18, 32)	387
quant_bn1 (QuantizeWrapperV2)	(None, 96, 18, 32)	129
quant_pool1 (QuantizeWrapperV2)	(None, 48, 9, 32)	1
quant_conv2 (QuantizeWrapperV2)	(None, 46, 7, 64)	18627

Table II: QAT Prepared Neural Network Architecture

Layer (type)	Output Shape	Param #
InputLayer	[-1, 98, 20, 1]	0
Conv2D	[-1, 96, 18, 32]	320 weights
Conv2D	[-1, 96, 18, 32]	+ 32 biases
MUL, ADD (BatchNorm)	[-1, 96, 18, 32]	64 scale
MUL, ADD (BatchNorm)	[-1, 96, 18, 32]	+ 32 offset
MaxPool2D	[-1, 48, 9, 32]	0

Table III: Quantized Model Architecture

Figure 6: Model architecture after NPU conversion [ 13 ] .

	Predicted Non-Marvin	Predicted Marvin
Real Non-Marvin	10746	64
Real Marvin	31	164

Table IV: Regular model confusion matrix

	Predicted Non-Marvin	Predicted Marvin
Real Non-Marvin	10501	309
Real Marvin	15	180

Table V: Quantized model confusion matrix

	Size [B]	Acc. [%]	i5 [ms]	MCXN947 [µs]
Regular	383674	99.14	50.67	-
Quantized	35744	97.06	0.42	228210
NPU	30576	97.06	-	3847

Table VI: Model statistics

NPU Acceleration

See all 8 papers

Sub-Model Short-Term Memory Convolutions for Keyword Spotting Systems on Device
Sep 28, 2026Paweł Warlewski, Artur Czeczko, Artur Szumaczuk +2Keyword Spotting Speech Processing
Samsung R&D Institute Poland · Samsung AI Center Warsaw
A Multiplication-Free Feature Extractor for Signal Classification: Keyword Spotting Case Study
Aug 17, 2026Radu Dogaru, Ioana DogaruSpeech Processing Tiny ML
Dept. of Applied Electronics and Information Engineering, National University of Science and Technology POLITEHNICA Bucharest, Romania
End-to-End Keyword Spotting on FPGA Using Graph Neural Networks with a Neuromorphic Auditory Sensor
May 10, 2026Wiktor Matykiewicz, Piotr Wzorek, Kamil Jeziorek +4Keyword Spotting Speech Processing
AGH University of Krakow, Poland · Embedded Vision Systems Group, Computer Vision Laboratory · Robotics and Technology of Computers Lab., ETSII, EPS, SCORE, I3US, Universidad de Sevilla

Sep 28, 2026J/K move · Enter open · S save

Sub-Model Short-Term Memory Convolutions for Keyword Spotting Systems on Device

Paweł Warlewski, Artur Czeczko, Artur Szumaczuk, Grzegorz Stefański+1

Samsung R&D Institute Poland · Samsung AI Center Warsaw

Keyword Spotting Speech Processing Convolutional Neural Networks On-Device Inference

Open paper AskPDF arXiv

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.

Abstract

NPU Acceleration

Explore similar work

Sub-Model Short-Term Memory Convolutions for Keyword Spotting Systems on Device

A Multiplication-Free Feature Extractor for Signal Classification: Keyword Spotting Case Study

End-to-End Keyword Spotting on FPGA Using Graph Neural Networks with a Neuromorphic Auditory Sensor