Smart Household Speech Recognition Based on Locally-Run Neural Network
Ivan Minárik, Matúš Vaño, Matúš Prekopa, Javlon Tursunov · 2024
This paper delves into speech recognition utilization in smart homes, focusing on privacy concerns with cloud-based solutions. We put STT services from Google, IBM and Microsoft against the locally-run solution based on OpenAI's Whisper to investigate their performance in real-world scenario. We consider latency, error rates, and environmental, speaker, and language dependencies on English and Slovak languages on our custom dataset. Our findings show Whisper's competitive performance, supporting minor languages like Slovak in smart home automation, while keeping the advantage of privacy.