I have a pre-processed collection of aligned audio-video clips and I want to push them through the AVHubert CoDA pipeline to obtain a clear Character Error Rate (CER) metric. The raw training and validation splits are ready; what’s missing is the glue code and know-how to hook my data into the official AVHubert / fairseq framework, run inference, and surface the CER report. Here’s what I expect at the end: • A working script or set of commands that load my dataset and execute AVHubert CoDA end-to-end (feature extraction, decoding, and CER calculation). • A brief README summarising environment setup (Python, PyTorch, fairseq, ffmpeg, KenLM, etc.) and any extra dependencies. • The final CER value plus the log files or JSON outputs that back it up, so I can replicate the run later. Acceptance criteria: running your README instructions on a fresh machine reproduces the same CER within tolerance. Everything will be shared through a private Git repo or secured cloud bucket—whichever is easiest for you.