Deploying speech enhancement directly on microcontroller units provides an energy-efficient, privacy-preserving front-end for voice-enabled edge computing. However, conventional deep neural architectures impose computational loads and memory footprints that exceed the strict resource boundaries of commodity microcontrollers. This paper introduces a Mel-scale STFT-domain TinyUNet comprising 13,651 parameters, optimized for on-device speech enhancement on an ESP32-S3 microcontroller operating at 240 MHz. Rather than processing full-resolution linear spectrograms, the architecture utilizes a 64-channel Mel filterbank representation, effectively halving the acoustic feature dimensionality while retaining the primary spectral envelope for causal mask estimation. Under full 8-bit integer (INT8) post-training quantization, the model occupies 45.57 KB of Flash memory and 18.06 KB of internal SRAM. Experimental evaluation on the DEMAND dataset across signal-to-noise ratios from -5 dB to +10 dB confirms that the quantized model achieves robust noise suppression, providing consistent improvements in objective quality and signal clarity over traditional statistical baselines with negligible performance loss relative to floating-point precision. Measured on-device execution of the complete pipeline, encompassing hardware-accelerated Fourier transforms, neural mask inference, and overlap-add synthesis, processes a 72 ms audio context window in 22.97 ms entirely within on-chip SRAM, establishing a practical baseline for deep neural speech enhancement on low-cost edge devices.

Speech enhancement,short-time Fourier transform (STFT),Microcontroller,Unet