A communication-assist device has been developed for stroke patients to clarify their unclear speeches. The patient talks as usual as the button is pressed,the converted voice plays immediately as the button is released. A custom system-on-a-chip (SoC) has been designedfabricated in 40nm CMOS, which has been integrated in the communication-assist device, which is 1,000× more energy-efficient than commercial application processors used in mobile phones.

Our team has developed personalized AI model for lightweight embedded implementations, which supports offline operations. Our intelligent audio SoC based on RISC-V boosts the AI performance by 12× while improving the energy efficiency by 100×. In addition, the novel zero-wait/zero skippingmulti-inference schemes further improve 3× computing time8× off-chip memory accesses respectively.

The “personalized AI model” for lightweight embedded implementation“ultra-low-power intelligent audio SoC” have been successfully applied to develop the communication-assist device for stroke patients. In addition, our team have also developed Bluetooth denoising microphoneslife period estimation for machinery tools based on same technologies. Future applications may include intelligent audio processing systemsdevices, low-power MCUIC designs.