Is this a subset relative to the original lists? How are the file paths mapped to the original collection?

#1
by arlofaria - opened

Thanks for doing the heavy lifting to prepare this dataset!

On an initial inspection, it seems like the data here on HF is significantly smaller than the amount originally published at https://github.com/facebookresearch/seamless_communication/blob/main/docs/m4t/seamless_align_README.md

For example, from the original TSV files for enA-ptA including the extension from 30 Nov 2023, I count over 6M audio segments totaling over 7K hours.

$ zcat seamless.dataset.metadata.public.enA-ptA.{,withduration.}tsv.gz | awk '{sum += ($4 - $3) / 16000} END {print sum / 3600}'
7354.71

However, the {src,tgt}.tar.gz files here appear to only have about 200K audio segments, totaling about 300 hrs. Is that right? Is this collection some filtered subset of the original?

Also, is there a way to map the dirname and filename paths from this collection to the metadata in the original collection's TSV files?

Any clarification would be much appreciated, thanks! :)

Center for Language and Speech Processing @ JHU org

Hi,

The audio files in the metadata are from online sources. A big issue we faced was that files were missing from their online sources or the urls may have changed. We did multiple download passes to make sure that we downloaded as many files as we could. We also used a duration check to verify if the files were similar. For some languages, the duration is greatly reduced as compared to what is present in the metadata.

The filenames correspond to the index number in the metadata 000/00/01 is 1.

Thanks for the clarification!

Did the duration check filter out a large proportion of the data? And was this a check on the metadata's reported duration of the entire audio (i.e. on the order of minutes)? Or rather a check that the translation-aligned segment pairs were of similar duration (i.e. on the order of seconds)?

BTW, I've just noticed that the "extension" metadata is sometimes a subset of the original metadata:
https://github.com/facebookresearch/seamless_communication/issues/467

So the corrected total duration for enA-ptA, ignoring the entirely duplicated segments in the "extension" metadata, is about 4.5K hours (not over 7K hours, as I originally posted).

Note that for other directions, such as enA-jaA, there is only "extension" metadata available -- that's not in the original metadata at all. Would JHU-CLSP plan to add these additional S2S language pairs in the future?

Sign up or log in to comment