A multilingual phoneme and model set: toward a universal base for automatic speech recognition

The amount of time, effort and expense that is required to incorporate a new language into an ASR system is extensive. It also is usually not possible to provide more than one language for speech recognition per system. The Core Technology Group of Conversant Voice Information Systems, Bell Laboratories has developed a multilingual phoneme and model set (MPMS) that is being used as a base for all telephone-based ASR continuous speech systems being developed in new languages. With a unified set such as the MPMS, it is possible that multiple languages could be available on one system. While the idea of a multilingual phoneme model set is not new, there has been no work that has used a large, telephone-based database consisting of continuous speech samples in more than two languages, obtains commercially acceptable word recognition rates, and that is ready to be marketed. Our system's phoneme set represents six different languages; we have built models based on three languages and tested them using two other languages (for which there were no models); and we have achieved very acceptable word recognition rates of better than 92% (field accuracy). These languages can be incorporated into an existing speech recognition system, available for customers.