- “I ripped all audio files from a game called Dead as Disco. It has FX, SFX. dialogue and some song stems. I think this could possibly help with datasets i think (also MX_Moonlight_152_VOX_Stem.ogg is a filtered thing in the game files for some reason)”

https://drive.google.com/drive/folders/1CDU0eqiOMFi4mmfX8vqx8VfVtHsWdjJQ?usp=sharing

Note: “It's very possible to have a dataset of sfx from tv shows and movies on Netflix - the 5.1 surround sound for movies/tv shows is different to that of Dolby Surround sound of music. Music has a lot of bleed and muddiness (artifacts) so that no one stem is soloed in a channel, and when it is, it's too muddy to be usable for anything, however in movies and TV shows, the surround sound is completely clean and has the sound effects pretty much soloed in a channel” - isling

Piano

“MAESTRO” is a dataset composed of about 200 hours of virtuosic piano performances captured with fine alignment (~3 ms) between note labels and audio waveforms.

https://magenta.tensorflow.org/datasets/maestro

GiantMIDI-Piano is a classical piano MIDI dataset contains 10,855 MIDI files of 2,786 composers. The curated subset by constraining composer surnames contains 7,236 MIDI files of 1,787 composers. GiantMIDI-Piano are transcribed from live recordings with a high-resolution piano transcription system

https://github.com/bytedance/GiantMIDI-Piano

Synth

- NSynth (Neural Synthesizer): Google Magenta, this foundational audio dataset features 305,979 four-second monophonic note snippets. It spans 1,006 commercial acoustic and electronic instruments across a full MIDI pitch range and multiple velocities, labeled by pitch, timbre, and envelope.

https://magenta.withgoogle.com/datasets/nsynth

- Slakh2100 (duplicate from the top): “TProvides 2,100 multi-track songs mixed from MIDI using professional sample-based synths. Every track gives you the full mix alongside perfectly isolated, bleed-free stems for synths, bass, drums, and guitars”.

https://zenodo.org/records/4599666

Drums

StemGMD: A Large-Scale Audio Dataset of Isolated Drum Stems for Deep Drums Demixing

(although drumsep used bigger dataset consisting of MIDI sounds to avoid bleeding, with XLN only)

Virtual drumkits

The “advantage is that you can have zero bleed between elements, which is not possible with real live drums.

You can create “more than 300 drumkits as virtual instruments (toontrack, kontakt, xln, slate, bfd, XLN ones are nice too (from their trigger and drums VST) + a Reaper framework to multiply that by 10 (using heavily different mixing processes for each drum elements), so potentially 3000 different sounding drumkits “

“one could use producer sample packs/kits for more modern samples” there are tons of packs around the net.

jarredou (rigo2):

“For those interested, I'm sharing on demand my drums separation dataset.

It's not a final version. I've realised after generating 130h of audio data that I've made a mistake in routing, leading to some occasional cowbell in snare stems. So it's [Kick/Snare-cowbell/Toms/HiHat/Ride/Crash] stems.

I've stopped it's rendering and will not make the final "mastering" stage that was planned.

I will make a clean no-cowbell version, but as I'm lacking free time, I don't know when, and as this one is here and already great sounding why not using it in the meantime.

Just don't mind the cowbell!”

Looks like it’s the thing:

~~http://rigaudio.fr/datasets/DrumsDataset.zip~~

Newer version in a better formatted version, with train/valid separated parts, generated mixtures and a fixed filename that was containing an extra space:

https://rigaudio.fr/datasets/DrumsDatasetv2.zip (25GB)

(still the same issues with some occasional cowbell in snare stems. “There are also few other percussions here and there on some little parts for some tracks (like tambourine in ride stem).”)

“I realise now that I totally forgot to lowercase all filenames before reuploading the dataset.

To avoid issues where some awaited filenames are hardcoded in ZFTurbo's script, the best way is so to lowercase all filenames in train/valid parts, convert the valid part to .wav files (no need for the train part that can handle flac correctly).

And lowercase the stem names in training part of the config file accordingly.”

“Can probably be useful to create electro drums separation dataset, free 50,000 drums MIDI files:” https://abasynthphony.gumroad.com/l/50000MIDIFilesforDanceMusicDrum?layout=profile

mesk’s metal drums dataset (drums in one stem):

??? (maybe ask him, looks I forgot pasting the link and can’t find it anymore)

Rhythm and lead guitar

https://www.mrtabs.com/

“He has isolated tracks for his videos that he makes, and it's free (or he does not know how to properly Patreon lock certain content on his website).

You can navigate to any tab page and look for the header: "Isolated TRACKS (mp3)" and find the textbox below where it says:

"Please sign up on Patreon, or if you are already a member, please login."

The last word links to a Patreon signup page, and if you sign in, it does not check if you are subscribed to his Patreon or not, it will give you access regardless.

Boom! Now you have access to 250+ Lead and rhythm guitar pairs. There is a goldmine worth of metal stuff in there too.

This is probably the closes we could ever get to having a contemporary rhythm/lead guitar dataset that is both relatively large, the rhythm and lead has their own tracks, its diverse, and actually includes songs that we like/listen to.

Only problem is all of them are exported in mp3 with a cutoff of 16khz, so it is equivalent 128 kbps, and the denoising that was done in post is pretty lazy.

However, I think if these parts are upscaled with FlashSR it would be great.

Or maybe Re-Amp a low pass filter version of the stems with the Ampltibe 5 presets that he also attaches to all tabs, and ensemble the remaining frequencies that way.

I personally would not recommend using the drum and bass stems, only the guitar parts, since the drum and bass are both programmed and are uniform.

The tone for every video is unique and tone matched to their respective albums and tracks. Even if it's not dead on, it's better than trying to use some yayhoos guitar doodles that uses the same amp/cabinet/simulator for every track.” Vinctekan

Ernhu

[China traditional music instrument dataset]

https://zenodo.org/records/8012071

EGFxSet: Electric guitar tones processed through real effects of distortion, modulation, delay and reverb

https://zenodo.org/records/7044411

Noise

http://research.spa.aalto.fi/publications/papers/icassp22-denoising/media/datasets/Gramophone_Record_Noise_Dataset.zip

_______

https://www.monotostereo.info/

“Helped me find not only tools but also other resources like research papers, etc on audio source separation in general. A fantastic resource for anyone into audio source separation”

For more links, check #resources and #datasets and Post dataset (you may encounter duplicates)

____