In-Vehicle Speech and Noise Corpora

Abstract

As in-vehicle speech systems become prevalent, there is a need for specific compilation of data in vehicle scenarios to develop/benchmark algorithms for speech systems. This paper describes the collection efforts and analysis of two corpora: (1) the UT-Dallas Vehicle Noise (UTD-VN) corpora and (2) the CU-Move in-car speech and noise corpora. The UTD-VN corpus is focused on addressing the variability of in-car noise environments. This corpus includes compilation of unique noise scenarios within the car (Engine idling, AC windows closed, etc.) as well as variability of these scenarios across different makes and models. Another aspect that in-car speech systems need to address along with noise is the emotional and task stress of the driver while performing the driving task. The CU-Move corpus focuses on collection of data to describe the variability of conversational speech in an in-car environment. A sample study is carried out where it is shown that these environments are unique across different vehicles using the UT-Dallas Vehicle Noise corpora. This shows that a detailed analysis of variability across vehicle platforms is necessary for successful deployment of speech systems. In our opinion, these corpora are the first to describe the environment variability along with conversational speech in an in-car environment.