First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Are Laboratory Databases Appropriate for Training and Testing Telephone Speech Recognizers?

Benjamin Chigier, Judith Spitz

NYNEX Artificial Intelligence Speech Technology Group, White Plains, NY, USA

Automatic speech recognition systems are typically trained on speech data collected in the laboratory and then tested on a mutually exclusive subset of the same data. Results of these tests may significantly overestimate performance in the field. It could be that systems should be trained and/or tested on spontaneously-produced real user field data. The goal of this study was to evaluate the performance of a speaker independent isolated word telephone network speech recognition system when tested on laboratory vs. real user data under two training scenarios: 1. trained on laboratory and 2. trained on real user data. The results of this experiment suggest that real user speech databases are needed to achieve high accuracy speech recognition results in the field. In addition, it appears that a system trained on user data can be accurately tested with either real user or laboratory speech databases.

Full Paper

Bibliographic reference.  Chigier, Benjamin / Spitz, Judith (1990): "Are laboratory databases appropriate for training and testing telephone speech recognizers?", In ICSLP-1990, 1017-1021.