Second International Conference on Spoken Language Processing (ICSLP'92)
Banff, Alberta, Canada
Speaker-independent speech recognition over the telephone network has begun to be practical for certain simple transactions such as automation of collect and third-number-billed telephone calls (Lennig, 1990). However, the current state of the art requires that such recognizers be trained on the specific vocabulary to be recognized and requires collecting training tokens of each vocabulary item from hundreds or often thousands of different speakers. The goal of the current study is to obviate the need for vocabulary-specific training, thus allowing a new vocabulary to be introduced merely by providing phonemic transcriptions for the words. In this paper we present speaker-independent results over the telephone network in which not only the speakers but also the vocabularies used in the training and test sets are disjunct. The vocabulary consists of the names of the 1,561 companies having common stock listed on the New York Stock Exchange in mid- 1986. Without the aid of a language model, we have achieved speaker-independent, vocabulary-independent recognition results of 96% correct.
Bibliographic reference. Lennig, Matthew / Sharp, Douglas / Kenny, Patrick / Gupta, Vishwa / Precoda, Kristin (1992): "Flexible vocabulary recognition of speech", In ICSLP-1992, 93-96.