Building a corpus and testing speech recognition

Speech recognition is still a rapidly developing area of research. It is used for all kinds of applications. In this article a corpus will be build that consists of sentences collected in human-computer dialogs. The corpus will be created in a Wizard-of-Oz environment where the user has to perform several tasks in a virtual environment related to navigation. The raw audio collected in these experiments was transcribed and then recognized by a large vocabulary recognizer. The first recognition results indicate that using a large vocabulary recognizer with a standard language and acoustic model gives poor recognition. It is curious that the recognition performance differs greatly with different participants in the experiment. Possible future steps are testing the corpus with adapted acoustic or language models, or using it for evaluating other speech recognition systems.