Considerable progress has been made in ASR technology in the recent years [Young, 1996] and many new products have been launched, but still completely new approaches need to be developed before robust, general-purpose speech recognizers will be available [Bourlard, 1995]. Current applications work only in relatively controlled environments and well-specified domains. For example, almost all applications are still restricted to the recognition of distinct words from a predefined vocabulary. Anyhow, such applications can be rather helpful in certain tasks.
By the vocabulary independent phoneme modeling as experimented in this work it is possible to recognize any spoken words, even those that have no meaning as long as they contain phonemes that can be trained for the models. The commercial production of the recognizers for unlimited vocabulary tasks have still been prevented by the excessive error rate and the need for careful speaker dependent training. However, some of the recognizer errors can be corrected by suitable post-processing of the phoneme sequences.