DeepMedia's DubSync: Bridging the INDOPAC Language Divide

In the face of China's expanding influence in the India-Pacific (INDOPAC) region, the U.S. Army faces an urgent need for precise translation capabilities, especially for INDOPAC's Low Resource Languages. This capability is crucial for analyzing publicly available information (PAI) to detect sentiment and indicators of information warfare (IW) and disinformation. Given the region's linguistic diversity, with languages like Hiligaynon, Bicolano, and Kapampangan, existing solutions fall short. Notably, the No Language Left Behind (NLLB) project has operational limitations, and Large Language Models (LLMs) such as GPT-4 are expensive and unsupportive of INDOPAC languages. DeepMedia aims to address these challenges by assessing the feasibility of enhancingĀits DubSync translation platform for specific Army applications through a Phase I SBIR. DubSync, with its vast dataset of 40 million sentence translations across 20 languages, offers enhanced translation accuracy for Army scenarios. It seamlessly integrates traditional NLP techniques with advanced models like GPT-J. Key system improvements include RNN-based sentiment analysis, a widened token window, and novel loss functions that enhance context transfer. The platform's innovative training strategy uses cyclic graph-based mappings for non-paired datasets. Notably, DubSync's CTNLLB model offers efficient Zero-Shot Translation for English to rare languages, eliminating the need for paired datasets. Furthermore, DubSync employs patented technologies in machine translation, voice recognition, and AI algorithms, enabling swift and precise translations across 200 languages. Its pioneering use of large-scale transformer models, coupled with traditional NLP techniques, results in a globally trusted translation platform with a Word Error Rate (WER) under 3%. By leveraging the foundational NLLB model and curating specialized datasets, DubSync is optimized for INDOPACOM RARE languages, capturing regional nuances and military jargon. The platform's unique architecture and extended token capacity ensure context-rich translations, making it a cost-effective, rapid, and accurate solution for the U.S. Army's linguistic challenges in the INDOPAC region.