Armenian.cc
SilviAvayan/Armenian.cc
speech-recognitionmachine-translationevaluation-harnesslow-resource-languagemultimodal-analysisllm-evaluationtiktok
Summary
A specialized evaluation harness for a pipeline that transcribes Armenian TikTok videos using ElevenLabs ASR and generates Russian glosses and translations via Gemini. The project focuses on rigorous measurement of ASR model performance (36 runs across 33 models) and the full pipeline's accuracy, uncertainty, and failure modes, rather than just demonstrating functionality. It includes a benchmark dataset, detailed analysis scripts, and results documentation.