The MUSDB18 is a dataset of 150 full lengths music tracks (~10h duration) of different genres along with their isolated drums, bass, vocals and others stems. The dataset is split into training and test sets with 100 and 50 songs, respectively. All signals are stereophonic and encoded at 44.1kHz. Source: https://sigsep.github.io/datasets/musdb.html#musdb18-compressed-stems Image Source: https://sigsep.github.io/datasets/musdb.html#musdb18-compressed-stems

MedleyDB, is a dataset of annotated, royalty-free multitrack recordings. It was curated primarily to support research on melody extraction. For each song melody f₀ annotations are provided as well as instrument activations for evaluating automatic instrument recognition. The original dataset consists of 122 multitrack songs out of which 108 include melody annotations. The songs in MedleyDB were obtained from the following sources: * Independent Artists (30 songs) * NYU's Dolan Recording Studio (32 songs) * Weathervane Music (25 songs) * Music Delta (35 songs) MedleyDB contains songs of a variety of musical genres: Singer/Songwriter, Classical, Rock, World/Folk, Fusion, Jazz, Pop, Musical Theatre, Rap. For each song three types of audio content are given: a mix, stems, and raw audio. All types of audio files are .wav files with a sample rate of 44.1 kHz and a bit depth of 16. Source: https://medleydb.weebly.com/ Image Source: https://medleydb.weebly.com/ Audio Source: https://zenodo.org/record/1438309

The Synthesized Lakh (Slakh) Dataset is a dataset for audio source separation that is synthesized from the Lakh MIDI Dataset v0.1 using professional-grade sample-based virtual instruments. This first release of Slakh, called Slakh2100, contains 2100 automatically mixed tracks and accompanying MIDI files synthesized using a professional-grade sampling engine. The tracks in Slakh2100 are split into training (1500 tracks), validation (375 tracks), and test (225 tracks) subsets, totaling 145 hours of mixtures. Source: http://www.slakh.com/ Image Source: http://www.slakh.com/ Audio Source: http://www.slakh.com/

MIR-1K (Multimedia Information Retrieval lab, 1000 song clips) is a dataset designed for singing voice separation. It contains: * 1000 song clips with the music accompaniment and the singing voice recorded as left and right channels, respectively, * Manual annotations of pitch contours in semitone, indices and types for unvoiced frames, lyrics, and vocal/non-vocal segments, * The speech recordings of the lyrics by the same person who sang the songs. The duration of each clip ranges from 4 to 13 seconds, and the total length of the dataset is 133 minutes. These clips are extracted from 110 karaoke songs which contain a mixture track and a music accompaniment track. These songs are freely selected from 5000 Chinese pop songs and sung by researchers from MIR lab (8 females and 11 males). Most of the singers are amateur and do not have professional music training. Source: https://sites.google.com/site/unvoicedsoundseparation/mir-1k Audio Source: https://sites.google.com/site/unvoicedsoundseparation/sounddemosforjournal

MUSDB18-HQ is a high-quality version of the MUSDB18 music tracks dataset. The high-quality dataset consists of the same 150 songs, but instead of MP4 files (compressed with Advanced Audio Coding encoder at 256kbps, with bandwidth limited to 16kHz), the songs are provided as raw WAV files. Image Source: https://sigsep.github.io/datasets/musdb.html

The CocoChorales Dataset CocoChorales is a dataset consisting of over 1400 hours of audio mixtures containing four-part chorales performed by 13 instruments, all synthesized with realistic-sounding generative models. CocoChorales contains mixes, sources, and MIDI data, as well as annotations for note expression (e.g., per-note volume and vibrato) and synthesis parameters (e.g., multi-f0). ## Dataset We created CocoChorales using two generative models produced by Magenta: Coconet and MIDI-DDSP. The dataset was created in two stages. First, we used a trained Coconet model to generate a large set of four-part chorales in the style of J.S. Bach. The output of this first stage is a set of note sequences, stored as MIDI, to which we assign a tempo and add random timing variations to each note (for added realism). In the second stage, we use MIDI-DDSP to synthesize these MIDI files into audio, resulting in audio clips that sound like the chorales were performed by live musicians. This MIDI-DDSP model was trained on URMP. We define a set of ensembles that consist of the following instruments, in Soprano, Alto, Tenor, Bass (SATB) order: <ul> <li><strong>String Ensemble</strong>: Violin 1, Violin 2, Viola, Cello.</li> <li><strong>Brass Ensemble</strong>: Trumpet, French Horn, Trombone, Tuba.</li> <li><strong>Woodwind Ensemble</strong>: Flute, Oboe, Clarinet, Bassoon.</li> <li><strong>Random Ensemble</strong>: Each SATB part is randomly assigned an instrument according to the following: <ul class="nested"> <li><em>Soprano</em>: Violin, Flute, Trumpet, Clarinet, Oboe.</li> <li><em>Alto</em>: Violin, Viola, Flute, Clarinet, Oboe, Saxophone, Trumpet, French Horn.</li> <li><em>Tenor</em>: Viola, Cello, Clarinet, Saxophone, Trombone, French Horn.</li> <li><em>Bass</em>: Cello, Double Bass, Bassoon, Tuba.</li> </ul> </li> </ul> Each instrument in the ensemble is synthesized separately, with annotations for the high-level expressions used for each note (e.g., vibrato, note volume, note brightness, etc; all expressions shown here, and more details in Sections 3.2 and B.3 of the MIDI-DDSP paper) as well as detailed low-level annotations for synthesis parameters (e.g., f₀’s, amplitudes of each harmonic, etc). Because the MIDI-DDSP model skews sharp, we randomly applied pitch augmentation to the f₀’s (see Figure 2, here) to . All four audio clips for each instrument in the ensemble are then mixed together to produce an example in the dataset. Because all of the data in CocoChorales originate from generative models, all of the annotations perfectly correspond to the audio data. All in all, the dataset contains 240,000 examples, 60,000 mixes from each one of the four ensemble types above. Each ensemble has its own train/validation/test split All of the audio is 16 kHz, 16-bit PCM data. Each example contains: <ul> <li>A mixture</li> <li>Source audio for all four instruments <ul class="nested"> <li>Gain applied to each source</li> </ul> </li> <li>MIDI with tempo and precise timing</li> <li>The name of the ensemble with instrument names</li> <li>Note expression annotations for every note: <ul class="nested"> <li>Volume, Volume Fluctuation, Volume Peak Position, Vibrato, Brightness, and Attack Noise used by MIDI-DDSP to synthesize every note (see Sections 3.2 and B.3 of the MIDI-DDSP paper for more details)</li> </ul> </li> <li>Synthesis parameters for every source (250 Hz): <ul class="nested"> <li>Fundamental frequency (f₀), amplitude, amplitude of all harmonics, filtered noise parameters</li> <li>Amount of pitch augmentation applied</li> </ul> </li> </ul> ## Further Details A detailed view of the contents of the CocoChorales dataset is provided at this link. # Download For download instructions, please see this github page. The compressed version of the full dataset is 2.9 Tb, and the uncompressed version is larger than 4 Tb. There is a "tiny" version for download as well. MD5 Hashes for all zipped files in the download are provided here. # License The CocoChorales dataset was made by Yusong Wu and is available under the Creative Commons Attribution 4.0 International (CC-BY 4.0). # How to Cite If you use CocoChorales in your work, we ask that you cite the following paper where it was introduced: Yusong Wu, Josh Gardner, Ethan Manilow, Ian Simon, Curtis Hawthorne, and Jesse Engel. “The Chamber Ensemble Generator: Limitless High-Quality MIR Data via Generative Modeling.” arXiv preprint, arXiv:2209.14458, 2022. You can also use the following bibtex entry: @article{wu2022chamber, title = {The Chamber Ensemble Generator: Limitless High-Quality MIR Data via Generative Modeling}, author = {Wu, Yusong and Gardner, Josh and Manilow, Ethan and Simon, Ian and Hawthorne, Curtis and Engel, Jesse}, journal={arXiv preprint arXiv:2209.14458}, year = {2022}, }

The MuseScore dataset is a collection of 344,166 audio and MIDI pairs downloaded from MuseScore website. The audio is usually synthesized by the MuseScore synthesizer. The audio clips have diverse musical genres and are about two mins long on average. Due to copyright issues the dataset is not publicly available, but can be collected and processed with the provided source code. Source: Multitask learning for frame-level instrument recognition Image Source: https://biboamy.github.io/instrument-demo/demo.html Audio Source: Somewhere over the rainbow

The SynthSOD dataset contains more than 47 hours of multitrack music obtained by synthesizing orchestra and ensemble pieces from the Symbolic Orchestral Database (SOD) using Spitfire BBC Symphony Orchestra Professional Library. To synthesize the MIDI files from the SOD, we needed to fix the original files into the General MIDI standard, select a subsect of files that fitted into our requirements (e.g., containing only instruments that we could synthesize), and develop a new system to generate musically-motivated random annotations about tempo, dynamic, and articulation.

This publicly available data is synthesised audio for woodwind quartets including renderings of each instrument in isolation. The data was created to be used as training data within Cadenza's second open machine learning challenge (CAD2) for the task on rebalancing classical music ensembles. The dataset is also intended for developing other music information retrieval (MIR) algorithms using machine learning. It was created because of the lack of large-scale datasets of classical woodwind music with separate audio for each instrument and permissive license for reuse. Music scores were selected from the OpenScore String Quartet corpus. These were rendered for two woodwind ensembles of (i) flute, oboe, clarinet and bassoon; and (ii) flute, oboe, alto saxophone and bassoon. This was done by a professional music producer using industry-standard software. Virtual instruments were used to create the audio for each instrument using software that interpreted expression markings in the score. Convolution reverberation was used to simulate a performance space and the ensembles mixed. The dataset consists of the audio and associated metadata Nineteen scores were randomly selected from the OpenScore String Quartet corpus. The synthesis of these as woodwind ensembles was performed by a sound engineering professional using professional software. The scores were loaded into the music notation software Steinberg's Dorico. The string parts were allocated to flute, oboe, clarinet (or alto saxophone) and bassoon. Virtual instruments were used to create the audio for each instrument: Miroslav Philharmonik 2 by IK Multimedia for the saxophone and Intimate Studio Winds by 8Dio for the other parts. The quartets were then imported into Avid's Pro Tools where they were mixed and reverberation added. Metadata were generated..