The published duration does not align with the json files

#3
by cooldog - opened

Hi I calculated the segment durations in json files and it seems quite different from the published figure. So the published figure is not valid hour?

Here is the duration in hour calculated from the json files:

language	train	dev
Vietnamese	4576.189	9.722
French	6490.673	13.085
English	8247.417	14.737
Spanish	10419.67	14.269
Portuguese	2758.729	8.143
Arabic	683.575	4.22
Thai	3349.19	9.026
Russian	2590.04	8.488
Korean	9624.534	17.585
German	5541.162	12.011
Dutch	3595.561	8.891
Japanese	1661.463	8.225
Hindi	4019.625	8.752
Indonesian	1684.05	8.218
Italian	4096.837	8.763

Besides, I also noticed that for some videos (perhaps not all), there are only very few lines of transcriptions in the json files while the speech in the video is actually pretty long. For example https://www.youtube.com/watch?v=_J0TYOu17uQ (JA_000000173 in Japanese/dev.json) has only about 50 seconds of segments transcribed in total while the video is actually 5 min long.

After the dataset was released, some videos were made private by their authors. To respect copyright, we have deleted those data.

Sign up or log in to comment