Integer-Only Approximated MFCC for Ultra-Low Power Audio NN Processing on Multi-Core MCUs
Marco Fariselli, Manuele Rusci, Joël Cambonie, Éric Flamand
Abstract
Given the recent advances in the design of efficient Deep Neural Networks (DNN) for tiny edge devices, the feature extraction frontend has become a computation bottleneck for enabling audio processing on low-end MicroController Units (MCUs). To address this challenge, this work presents novel hardware-aware integer quantization schemes for the MelFrequency Cepstral Coefficients (MFCC) feature extractor. Our high-precision integer-only 32 bit approximated flow does not lead to accuracy degradation with respect to a full-precision implementation when feeding multiple DNN models for Audio Keyword Spotting applications. In contrast, a second low-precision 16-bit approximated MFCC algorithm presents a 0.6% lower accuracy but results $3\times$ faster. Additionally, by leveraging on an 8-cores MCU, GAP8, our solution results $9.8\times$ faster than the full precision MFCC deployed on an FPU-suited MCU. When integrated within an optimized end-to-end system for Keyword Spotting, a GAP8-based audio smart device presents an overall power consumption as low as 3.4mW, demonstrating up to 35 days of lifetime with a single AA battery.