Enhanced Language Translation of Spoken Text and AI Cross-Regional Model Training

This effort primarily addresses three USFJ requirements: A) Joint, cross-regional AI model training capabilities: Artificial Intelligence technologies operate on a reverse depreciation curve--additional use makes the software smarter and more effective.  USFJ seeks to stand up a new AI training concept through this project by enabling multiple USFJ components at different geographic locations and across the services to participate in shared, collaborative AI model training efforts.  B) Maximum-accuracy foreign language translation: In operational and industrial shipyard contexts, mistranslations can lead to loss of life or severe property damage.  USFJ requires language translation models that can translate military-domain content with the greatest possible clarity and accuracy. C) Spoken text translation capability: In both a) operational/J3 use cases [e.g. translating recorded military briefings from English to Japanese] and b) intelligence/J2 use cases [e.g. translating Russian/Chinese audio/video files], multiple USFJ components require enhanced capability to translate spoken text.  In response to these requirements, USFJ has two primary objectives for this effort: USFJ Objective #1 - Enhanced Capability for Spoken Text Translation:  Milestone 0003 of the parent Defense Language Institute (DLI) Language Proficiency Assessment Directorate (LPAD) Phase II SBIR project (AF221-DCSO1) involved a highly manual, command line process for integrating Automatic Speech Recognition (ASR) and MT technologies.  While this solution accomplished LPAD’s objective of generating English transcripts for recorded foreign-language Oral Proficiency Interviews, it is not usable in practical settings by DoD personnel who do not have specialized computer software training.  This Navy Subsequent Phase II project will conduct the additional required research, prototyping, and testing work needed to make the functionality developed by LPAD available to DoD users at scale, extending the benefit of DLI LPAD’s innovation to multiple new commands and use cases.  USFJ Objective #2 - Experiment with Processes for Joint, Cross-regional AI Model Fine-tuning:  Milestone 0004 of LPAD’s Phase II SBIR project involved fine-tuning the foreign language MT model to increase accuracy scores.  That work demonstrated significant capability gains, but it was conducted a) by a single service [USAF] b) on a single installation [Presidio of Monterey] c) on a narrow, focused set of content [OPIs].  This USFJ innovation and research effort seeks to expand that work to include AI model training conducted a) across the military services, b) from different USFJ installations, and c) across diverse content types.  To accomplish this objective, translators from the five participating USFJ commands will participate in a joint model training experiment--they will a) generate training data as they translate mission-related content and b) synthesize the training data to optimize model performance.