Automatic Spoken Language Recognition for Machine Foreign Language Translation (MFLT)

Voci Technologies Incorporated (Voci) is partnering with Bhiksha Raj, a Voci advisor and Professor at Carnegie Mellon University, to prototype an Automated Spoken Language Recognition System (ASLRS) specifically designed to enhance the utility of Speech to Speech Machine Foreign Language Translation (MFLT) systems for the warfighter. The proposed ASLRS will be developed by tailoring Voci"s existing language identification technologies to fulfill the requirements of an efficient MFLT preprocessor. To provide best in class accuracy, a combination of techniques will be used including acoustic UBM-GMM (Universal-Background-Model/Gaussian-Mixture-Model), and PPRLM-SVM (Parallel-Phonetic-Recognition-Language-Model/Support-Vector-Machine). To meet the ASLRS real-time requirements a groundbreaking, patent pending, multi-language phonetic dictionary capable of doing phonetic recognition in all 6 target languages in a single pass will be utilized. An open-set based solution will be provided so that the ASLRS will recognize when an out of domain language is spoken. To ensure that the resulting ASLRS is generally applicable, it will be architected to be an open system that is easily integratable with existing MFLT solutions. To ensure that the system provides reliable results, even in noisy environment, the system will incorporate noise robust features. Finally, to address the shortcomings of existing solutions in real-world field conditions, the Voci Team will integrate an online incremental learning capability into the ASLRS so that it can adapt to different accents and noise conditions that exist during field use. At the end of Phase I the Team will demonstrate the prototype ASLR system. We believe the final implementation will yield a revolutionary new ASLR capability.