laion-ai / audio-dataset Goto Github PK

@marianna13 Hi Mariana,
It seems like there are several options to have a Music Dataset. However, could you recommend me one (or many) for training the MuLaN model?

They used 44 million music recordings (almost 370K hours). The following table show some examples of their texts of 3 different types.

https://github.com/speechcolab/gigaspeech

Assigned: Knoriy

https://zenodo.org/record/1188976

https://www.juno.co.uk/products/wu-tang-clan-the-charmels-cream/861381-01/

https://research.google.com/audioset/

Where do we get the following files from?
Any help would be appreciated.
Thanks in advance.

https://audiocaps.github.io/

Hi, can you share some ways to download Freesound? e.g. How to use Linux scripts to download these audio.

https://github.com/facebookresearch/covost

https://github.com/karolpiczak/esc-50

Assigned: Knoriy

https://github.com/CheyneyComputerScience/CREMA-D

https://www.twine.net/blog/100-audio-and-video-datasets/?utm_source=sales&utm_medium=email&utm_campaign=zapent1&utm_content=atest

Current location AWS S3 bucket, not yet prepared

https://annotator.freesound.org/fsd/release/fsd50k/

https://zenodo.org/record/4783391#.ygdaa9-znpy

Most of the urls in the Freesound (no overlap)train+test.csv files are invalid. When I visit the url, I find the result like this:

https://freesound.org/apiv2/sounds/621393/download/?format=api

How can I download the dataset correctly? Thank you!

This repo is great. I always wanted to benchmark webdataset for audio. A couple of questions:

1. did you find flac to be a good trade-off between decoding performance and file-size? have you tried mp3 instead?
2. did you benchmark the pipeline against plain torch.data with torchaudio or the new torch data pipes? Maybe adding the benchmark to https://github.com/faroit/python_audio_loading_benchmark/ to give this a go?
3. How is partial decoding seeking be typically done with webdatasets, when storing long audio but at decoding stage, only random chunks are being read. Is seeking supported? If yes, does this slow down the i/o pipeline?

https://openslr.org/115/

https://sail.usc.edu/iemocap/

Hi,
Thanks for sharing the wonderful code.

According to the readme of data preprocess (here)
there should be a key of 'tag' (containing labels) in the output JSON file after preprocessing.

This tag extraction/creation is missing in the preprocess_FSD50K.py file.

Am I understanding something incorrectly or there is 'tag' creation missing in the file?

Thanks,
Saksham

https://keithito.com/lj-speech-dataset/

https://paperswithcode.com/dataset/librispeech

@rvencu @rom1504
We need more data in the next step. The data we need in the ranking of priority is:

1. Audio data with natural text description(s).
2. Audio data with other labels, and "made up" a text description for the audio.

For audio data with natural text description, we further need:

• MACS - Multi-Annotator Captioned Soundscapes: a dataset containing audio captions and corresponding audio tags for a number of 3930 audio files of the TAU Urban Acoustic Scenes 2019 development dataset (airport, public square, and park). The files were annotated using a web-based tool.
• Free Sound: scrape audio and text description from Free Sound. It is ok that the texts are a bit noisy.
• High-quality sound effect library with similar quality as BBC Sound Effect: such as https://www.sound-ideas.com/Default.aspx or https://www.boomlibrary.com/ who has high-quality text descriptions of the audio rather than tags and labels.
• Music review websites: such as Pitch Fork

For audio data with other labels, we need to collect new large datasets while converting our current dataset with tag labels.

The datasets in top priority are those with large size and easy to turn labels into a text description:

(The following datasets all are those with tag labels of the audio)

• AudioSet
• ACAV100M
• UrbanSound8K
• FSD50K

The datasets we currently have that need converting labels to text are:

• Sonniss Game effectsL: https://deploy.laion.ai/0fed69941baaabaeccedc2aaaaaaaaaa/
• WeSoundEffects: https://deploy.laion.ai/0fed69941baaabaeccedc2aaaaaaaaaa/
• Paramount Motion - Odeon Cinematic Sound Effects: https://deploy.laion.ai/0fed69941baaabaeccedc2aaaaaaaaaa/
• Free to use sounds: https://deploy.laion.ai/0fed69941baaabaeccedc2aaaaaaaaab/

We should come up with a unified way of converting tags to text. We could reference how CLIP did that (in converting classification to natural text).

current location on AWS S3 bucket

https://urbansounddataset.weebly.com/urbansound8k.html

https://academictorrents.com/details/d2b2ae5e3ec4fd475d6e4c517d4c8752a7aa8455

https://freesound.org/

https://www.mllp.upv.es/europarl-st/

https://www.epidemicsound.com/

How can we get the dataset you are using, I can not rerun the preprocessing code for the dataset downloaded from their website

https://mlcommons.org/en/peoples-speech/

Hi, I want to ask how to download Epidemic Sound dataset. Following the link in the audiostock csv file, e.g. https://audiostock.net/audio/1111457/play , I find that I can download it by Windows browser, but it cannot used on linux system.

https://mlcommons.org/en/multilingual-spoken-words/

Similarly to image datasets, it's better to first save a url + metadata file as parquet
That can be distributed without copyright issue

Then a tool like img2dataset can handle the download

Let's add that in the readme here

https://sound-effects.bbcrewind.co.uk/

https://www.paramountmotion.com/odeon-sound-effects

https://github.com/mdeff/fma, and also https://academictorrents.com/details/dba20c45d4d6fa6453a4e99d2f8a4817893cfb94

https://zenodo.org/record/5114771#.yq4kbnbmlb1

Congratulations for executing the herculean effort of putting together this dataset!
Where can one find the access information for the data in s3://s-laion-audio/?