About the dataset

[yudongliu97]

Thank you for the release of the dataset! Can you show me an example how to get the youtube link given a youtube ID in your dataset? Such as oPRC3zComfA_000117.

[ac-alpha]

@linyifan123456 I am not an author of this paper, but based on the dataset samples provided on the dataset website, the youtube_id can be split into two parts separated by the underscore (_) character.

• The first part is the YouTube URL of the source video (You can create the original YouTube URL by using this format https://www.youtube.com/watch?v=<youtube_url_extracted>).
• The second part is the start time (in seconds) in the original YouTube URL from where the dataset sample is constructed.

Example - oPRC3zComfA_000117 can be split into oPRC3zComfA giving us the source video URL as https://www.youtube.com/watch?v=oPRC3zComfA and the start time in the video as 117 seconds which is 1:57 timestamp.

However, some samples do not have the second part (start time) which is strange. I will request the authors to clarify more on that.

[LoieSun]

Thank you for your interest in our dataset. As @ac-alpha mentioned, the number following the underscore indicates the video's start time. Our dataset is based on AudioSet and VGGSound; the samples marked with a start time are sourced from VGGSound, whereas the remaining samples are from AudioSet. For additional download options, please refer to the homepages of AudioSet and the VGGSound.

[Sreyan88]

Hi @LoieSun ,

I tried matching the names in the dataset you provided with the file names from the original AudioSet, but I found only 53k matches. Am I missing something here?

[LoieSun]

hi @Sreyan88 , I would like to know more details about this problem if possible. It is worth noting that in Auto-ACD, the filenames of samples sourced from AudioSet are not marked with a start time.

[JunZhan2000]

hi @Sreyan88 , I would like to know more details about this problem if possible. It is worth noting that in Auto-ACD, the filenames of samples sourced from AudioSet are not marked with a start time.

Hello, very good work, I want to use this dataset in my work! I would like to ask if I want to download the dataset, do I have to crawl the data from youtube to align? If you can, can you provide a direct download link?

[LoieSun]

@junzhan18 hi, thank you for your interest. For text, you could download from this link. For audio, you could download the original AudioSet and VGGsound, and separate audio from video.

[Sreyan88]

Hi @LoieSun , Thank You so much for your reply. I am trying to match with these filenames but only able to achieve 53k matches. Any help would be highly appreciated, and excellent work again!

[JunZhan2000]

@junzhan18 hi, thank you for your interest. For text, you could download from this link. For audio, you could download the original AudioSet and VGGsound, and separate audio from video.

Audioset only provides YouTube links and extracted features. If I need the original audio, do I still need to crawl it myself?

[JunZhan2000]

@junzhan18 hi, thank you for your interest. For text, you could download from this link. For audio, you could download the original AudioSet and VGGsound, and separate audio from video.

Hello, I am downloading this dataset, but I also encountered a problem that does not match Audioset. If it is convenient, can you add a WeChat or email exchange? I cannot find your email address.

[JunZhan2000]

Hi @LoieSun , Thank You so much for your reply. I am trying to match with these filenames but only able to achieve 53k matches. Any help would be highly appreciated, and excellent work again!

I downloaded auto-cad from huggingface, http://huggingface.co/datasets/Loie/Auto-ACD/viewer/default/train? p=19218, downloaded audioset metadata from this link http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/unbalanced_train_segments.csv, but does not match their youtube id completely.

[Sreyan88]

Hi @junzhan18 , I am facing the same problem. Please lmk if you were able to solve it.

[JunZhan2000]

Hi @LoieSun , Thank You so much for your reply. I am trying to match with these filenames but only able to achieve 53k matches. Any help would be highly appreciated, and excellent work again!

I downloaded auto-cad from huggingface, http://huggingface.co/datasets/Loie/Auto-ACD/viewer/default/train? p=19218, downloaded audioset metadata from this link http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/unbalanced_train_segments.csv, but does not match their youtube id completely.

@LoieSun VGGSound's data matches, but AudioSet's doesn't.

[LoieSun]

@junzhan18 @Sreyan88 hi, for data from AudioSet, we named them using a template, 'Y+youtube id'. For example, 'Yae9Icdo66YA' indicates the YouTube video with id 'ae9Icdo66YA', and you could find it in the _unbalanced_trainsegments.csv.

misc: -my WeChat: Loie142857 -E-mail address: loiesun411@gmail.com

[Sreyan88]

Hi @LoieSun ,

Thank You for your response. However, I still don't find a match. For example, for YcxmYcIGBZDs (in auto-acd) I don't find cxmYcIGBZDs in unbalanced_train_segments.csv.

Am I missing something?

[Sreyan88]

for your above given example, I dont even find 'ae9Icdo66YA' in unbalanced_train_segments.csv!

[JunZhan2000]

Hi @LoieSun ,

Thank You for your response. However, I still don't find a match. For example, for YcxmYcIGBZDs (in auto-acd) I don't find cxmYcIGBZDs in unbalanced_train_segments.csv.

Am I missing something?

hi, you can use this code to load audioset metadata, it's ok

import pandas as pd audioset_metadata = "path/to/unbalanced_train_segments.csv" audioset = pd.read_csv(audioset_metadata, usecols=[0])