In automatic speech recognition applications, recognition rates are usually estimated using clean laboratory speech. This kind of data is of course needed to train the system before it is put in service, but it does not reflect the true behavior of users. As a result, the actual recognition rates are much lower. In addition, a good part of the input data is not even valid, not being among the vocabulary words. We present a short description of two field databases containing close to 26,000 speech tokens from real users. The field databases were extracted from two voice response systems equipped with isolated word recognition, operating on the French telephone network.
Bibliographic reference. Mathan, Luc / Morin, Dominique (1991): "Speech field databases: development and analysis", In EUROSPEECH-1991, 509-512.