IoT Device Control with Offline Automatic Speech Recognition on Edge Device

Abstract

Automatic speech recognition (ASR) on edge device still barely used in industry. Most of ASR such as speech-to-text commonly depend on the network presence. This is the disadvantages of using the server-based architecture that cannot always be reliable and available. In some conditions, the speech recognition need to be used in offline condition and edge-processing. Raspberry Pi, one of many edge devices will use offline ASR platform with only runs on python language. These platforms will encourage to do speech recognition in offline condition and edge device processing. Vosk, Picovoice/Cheetah, Sopare runs with python language and some of it required Word-Error-Rate (WER) on the result to validate result value. The 5 simple commands trained on each platform. Overall result, these 3 platforms are working well in offline environment with good accuracy (>80%). Vosk has around 84% accuracy, Picovoice/Cheetah has around 72% and Sopare has around 95% accuracy. Picovoice/Cheetah need an internet connection at first application startup.