Three approaches to multilingual phone recognition

This paper investigates and compares three different approaches of multilingual phone recognition (MPR). Two types of MPR approach are defined according to the language identification (LID) process of the system: explicit-LID where language identification is mandatory, and implicit-LID where LID is an integrated part of the MPR process. The OGI-TS database is employed to perform the isolated and continuous MPR experiments. Three of the world's most spoken languages, English, Mandarin and Spanish, are selected as the target languages for the system. Experimental results indicate that different MPR approaches should be employed for different applications according to the degree of LID accuracy that can be achieved from the input test utterance. If high LID accuracy is achievable, the MPR approach that depends on LID can obtain better performance. Conversely, the implicit-LID MPR approach is more appropriate.