Recently, interest in voice assistants and question-and-answer systems has increased. Automotive OEMs are also pursuing the commercialization of voice assistant functions by introducing external IT technology. However, ICT companies are limited in developing vehicle-specific services because they perform well in a universal domain rather than vehicle-specific functions. When developing a vehicle-specific voice recognition system, it is difficult to provide services provided by ICT companies together due to content supply and demand problems. To solve this problem, we tried combining vehicle-specific voice recognition with ICT companies' voice recognition. However, in the presence of a multi-voice recognition system, we cannot determine which system's result should be finally selected. This paper proposes a novel method for seamlessly combining various voice recognition systems so that both vehicle-specific and ICT companies provide services and vehicle-specific voice recognition system. The vehicle-specific voice recognition system consists of four steps: First, the intention analyzer classifies a domain of user’s utterance, extracts slots of named entities. Second, the domain classifier distinguishes whether it is an internal domain of a vehicle or an out-of-domain. Third, a state-of-art deep learning confidence controller is applied that can be determined as out-of-domain for unseen data. Finally, the final decision module determines which voice recognition system to use. The classification performance of multiple voice recognition systems could be improved to 97.6% by applying the proposed method compared to 85.2% of the existing deep learning algorithm in Hyundai NLU2021 database.

Hyundai Motor Company: Sungsoo Park, Changwoo Chun, Jihoon Kim