An 8.62μW 75dB-DR_SoCEnd-to-End Spoken-Language-Understanding SoC with Channel-Level AGC and Temporal-Sparsity-Aware Streaming-Mode RNN

Voice-controlled IoT nodes and wearable devices require integrated real-time ultra-low-power audio classification circuits to perform tasks such as Keyword Spotting (KWS) and Spoken Language Understanding (SLU). In conventional ADC+DSP implementations [1]-[2], the analog front-end (AFE) and digital feature extractor (FEx) together accounted for >50% of the system power. Analog FEx [3]-[9] reduces power by direct analog-to-feature conversion. Voltage-domain FEx [3]-[6] achieved <0.5μW power but only demonstrated <6 classes KWS. Time-domain FEx [7]-[9] achieved 86%-to-91.5% KWS accuracy with 10-to-12 classes but needed amplitude-normalized input or a costly off-chip classifier. In addition, prior designs [1]-[10] were limited to single-word audio inputs and did not consider continuous speech inputs required by SLU. Real-world operation also requires >60dB input range to cope with the variation of speech volume [11] and the speaker distance from the microphone.