MCAT: Scaling Many-to-Many Speech-to-Text Translation with MLLMs to 70 Languages

MCAT addresses two critical bottlenecks in current MLLM-based speech translation: the heavy English-centric bias of datasets and the computational cost of processing long speech token sequences. While the project shows technical depth in scaling to 70 languages and optimizing inference speed, it faces extreme frontier risk. Major labs like Meta (with SeamlessM4T) and OpenAI (with Whisper and GPT-4o's native audio capabilities) are already dominating the multilingual speech space. The '10 forks vs 0 stars' signal indicates immediate interest from researchers and engineers who are likely pulling the code to replicate results or benchmark against internal models, but the lack of stars suggests it hasn't yet built a community. The defensibility is low because the architectural innovations in MLLM-speech integration (like token compression or cross-modal attention) are being commoditized rapidly. A frontier lab could integrate these specific 70-language optimizations into a general-purpose model in a single training cycle if they found the approach superior to current methods.