MusicNet-16k + EM for YourMT3

< UNDER CONSTRUCTION >

About this version:

This particular variant of the MusicNet dataset has been resampled to a 16 kHz-mono-16-bit-wav format, which makes it more suitable for certain audio processing tasks, particularly those that require lower sampling rates. We redistribute this data as a part of YourMT3 project. The license for redistribution is attached.

Moreover, this version of the dataset includes various split options derived from previous works on automatic music transcription as python dictionary (see README.md). Below is a brief description of available split options:

MUSICNET_SPLIT_INFO = {
    'train_mt3': [], # the first 300 songs are synth dataset, while the remaining 300 songs are acoustic dataset. 
    'train_mt3_synth' : [], # Note: this is not the synthetic dataset of EM (MIDI Pop 80K) nor pitch-augmented. Just recording of MusicNet MIDI, split by MT3 author's split. But not sure if they used this (maybe not).
    'train_mt3_acoustic': [],
    'validation_mt3': [1733, 1765, 1790, 1818, 2160, 2198, 2289, 2300, 2308, 2315, 2336, 2466, 2477, 2504, 2611],
    'validation_mt3_synth': [1733, 1765, 1790, 1818, 2160, 2198, 2289, 2300, 2308, 2315, 2336, 2466, 2477, 2504, 2611],
    'validation_mt3_acoustic': [1733, 1765, 1790, 1818, 2160, 2198, 2289, 2300, 2308, 2315, 2336, 2466, 2477, 2504, 2611],
    'test_mt3_acoustic': [1729, 1776, 1813, 1893, 2118, 2186, 2296, 2431, 2432, 2487, 2497, 2501, 2507, 2537, 2621],
    'train_thickstun': [], # the first 327 songs are synth dataset, while the remaining 327 songs are acoustic dataset.  
    'test_thickstun': [1819, 2303, 2382],
    'train_mt3_em': [], # 293 tracks. MT3 train set - 7 missing tracks[2194, 2211, 2227, 2230, 2292, 2305, 2310], ours
    'validation_mt3_em': [1733, 1765, 1790, 1818, 2160, 2198, 2289, 2300, 2308, 2315, 2336, 2466, 2477, 2504, 2611], # ours
    'test_mt3_em': [1729, 1776, 1813, 1893, 2118, 2186, 2296, 2431, 2432, 2487, 2497, 2501, 2507, 2537, 2621], # ours
    'train_em_table2' : [], # 317 tracks. Whole set - 7 missing tracks[2194, 2211, 2227, 2230, 2292, 2305, 2310] - 6 test_em
    'test_em_table2' : [2191, 2628, 2106, 2298, 1819, 2416], # strings and winds from Cheuk's split, using EM annotations
    'test_cheuk_table2' : [2191, 2628, 2106, 2298, 1819, 2416], # strings and winds from Cheuk's split, using Thickstun's annotations
}

About MusicNet:

The MusicNet dataset, originally released in 2016 by Thickstun et al., "Learning Features of Music from Scratch". It is a collection of music recordings annotated with labels for various tasks, such as automatic music transcription, instrument recognition, and genre classification. The original dataset contains over 330 hours of audio, sourced from various public domain recordings of classical music, and is labeled with instrument activations and note-wise annotations.

About MusicNet EM:

MusicNetEM are refined labels for the MusicNet dataset, in the form of MIDI files. They are aligned with the recordings, with onset timing within 32ms. They were created using an EM process, similar to the one described in the Ben Maman and Amit H. Bermano, "Unaligned Supervision for Automatic Music Transcription in The Wild". Their split (Table 2 of this paper) derived from another paper, Kin Wai Cheuk et al., "ReconVAT: A Semi-Supervised Automatic Music Transcription Framework for Low-Resource Real-World Data".

License:

CC-BY-4.0