A 34.7 µW Speech Keyword Spotting IC Based on Subband Energy Feature Extraction

In the era of Internet Things (IoT), voice control has enhanced human–machine interaction and accuracy keyword spotting (KWS) algorithms reached 97%; however, high power consumption KWS caused by their huge computing storage requirements limited application in Artificial Intelligence (AIoT) devices. this study, features are extracted utilizing fast discrete cosine transform (FDCT) for frequency-domain transformation to shorten process calculating logarithmic spectrum cepstrum. The designed system is a two-stage wake-up system, with sound detection (SD) awakening KWS. inference network achieved using time-division computation, reducing clock an ultra-low frequency 24 kHz.At same time, implementation depthwise separable convolution neural (DSCNN) greatly reduces parameter quantity computation. Under GSMC 0.11 µm technology, post-layout simulation results show that total synthesized area entire circuit 0.58 mm2, 34.7 µW, F1-score 0.89 10 dB noise, which makes it suitable as AIoT