Abstract
The Australian In-Car Speech Corpus is a multi-channel recording of a series of prompts from an in-car navigation task collected over a range of speakers in a variety of driving conditions. Its purpose is to provide a significant resource of speech data appropriate for investigating speech processing needs in the adverse environment of a car. Utterances spoken by 50 speakers were collected in seven different driving conditions, providing the foundation for investigation into noisy, speaker-independent speech processing. Speech recognition experiments are performed to validate the data, to provide baseline results for in-car speech recognition research, and to show that this data can improve speech recognition performance under adverse in-car conditions for Australian English when adapting from American English acoustic models.