WMT24 Metrics Task Speech Subset

[jack89roberts]

The WMT 2024 metrics task had a new speech subset. You can get all the WMT metrics evaluation data using this repo: https://github.com/google-research/mt-metrics-eval (has source text, human translation, machine translations for many models, human ratings, and metric ratings). We should be able to pretty easily compute results for the speech subset and see how metrics compare between them, which I don't think WMT included in their results paper. They haven't released audio as far as I can tell, but I dropped the organisers an email to see if it's available, but either way we could also add BLASER with text at least.

There are some other subsets that might be interesting too, e.g. a social text one, plus even a separate text chat translation challenge I hadn't noticed.

[jack89roberts]

Audio is available here but may take some time to match up with the relevant lines in the metrics data: https://data.statmt.org/wmt24/general-mt/wmt24_GeneralMT-audio.zip