DRUM TRANSCRIPTION VIA JOINT BEAT AND DRUM MODELING USING - - PowerPoint PPT Presentation

▶
drum transcription via joint beat and drum modeling using
SMART_READER_LITE
LIVE PREVIEW

DRUM TRANSCRIPTION VIA JOINT BEAT AND DRUM MODELING USING - - PowerPoint PPT Presentation

Dec 02, 2022 329 likes •1.14k views

DRUM TRANSCRIPTION VIA JOINT BEAT AND DRUM MODELING USING CONVOLUTIONAL RNNs Richard Vogl 1,2 , Matthias Dorfer 2 , Gerhard Widmer 2 , Peter Knees 1 richard.vogl@tuwien.ac.at, matthias.dorfer@jku.at, gerhard.widmer@jku.at,

slide-1
SLIDE 1

Richard Vogl1,2, Matthias Dorfer2, Gerhard Widmer2, Peter Knees1

richard.vogl@tuwien.ac.at, matthias.dorfer@jku.at, gerhard.widmer@jku.at, peter.knees@tuwien.ac.at

DRUM TRANSCRIPTION VIA 
 JOINT BEAT AND DRUM MODELING USING CONVOLUTIONAL RNNs

1 2
slide-2
SLIDE 2

WHAT IS DRUM TRANSCRIPTION?

2

Input: western popular music containing drums Output: symbolic representation of notes played by drum instruments

slide-3
SLIDE 3

WHAT IS DRUM TRANSCRIPTION?

Focus on the three major drum instruments:

  • bass or kick drum (KD)
  • snare drum (SD)
  • hi-hat (HH)

Reasons:

  • Dominant instruments: most onsets
  • Common subset for public datasets
3

KD SD HH

slide-4
SLIDE 4

SYSTEM OVERVIEW

4 signal preprocessing NN 
 feature extraction 
 event detection classification peak picking NN training audio events
slide-5
SLIDE 5

SYSTEM OVERVIEW

4 signal preprocessing NN 
 feature extraction 
 event detection classification peak picking NN training audio events spectrogram t [s] f [Hz]
slide-6
SLIDE 6

SYSTEM OVERVIEW

4 signal preprocessing NN 
 feature extraction 
 event detection classification peak picking NN training audio events spectrogram t [s] f [Hz] t [s] activation functions
slide-7
SLIDE 7

SYSTEM OVERVIEW

4 signal preprocessing NN 
 feature extraction 
 event detection classification peak picking NN training audio events spectrogram t [s] f [Hz] t [s] activation functions
slide-8
SLIDE 8

ISSUES OF CURRENT SYSTEMS

5
slide-9
SLIDE 9

Performance not satisfying on real music

ISSUES OF CURRENT SYSTEMS

5
slide-10
SLIDE 10

Performance not satisfying on real music Do not produce additional information for transcripts
 drum onset detection vs drum transcription

ISSUES OF CURRENT SYSTEMS

5
slide-11
SLIDE 11

Performance not satisfying on real music Do not produce additional information for transcripts
 drum onset detection vs drum transcription

  • bars lines

ISSUES OF CURRENT SYSTEMS

5
slide-12
SLIDE 12

Performance not satisfying on real music Do not produce additional information for transcripts
 drum onset detection vs drum transcription

  • bars lines
  • tempo

ISSUES OF CURRENT SYSTEMS

5
slide-13
SLIDE 13

Performance not satisfying on real music Do not produce additional information for transcripts
 drum onset detection vs drum transcription

  • bars lines
  • tempo
  • meter

ISSUES OF CURRENT SYSTEMS

5
slide-14
SLIDE 14

Performance not satisfying on real music Do not produce additional information for transcripts
 drum onset detection vs drum transcription

  • bars lines
  • tempo
  • meter
  • dynamics / accents

ISSUES OF CURRENT SYSTEMS

5
slide-15
SLIDE 15

Performance not satisfying on real music Do not produce additional information for transcripts
 drum onset detection vs drum transcription

  • bars lines
  • tempo
  • meter
  • dynamics / accents
  • stroke / playing technique

ISSUES OF CURRENT SYSTEMS

5
slide-16
SLIDE 16

Performance not satisfying on real music Do not produce additional information for transcripts
 drum onset detection vs drum transcription

  • bars lines
  • tempo
  • meter
  • dynamics / accents
  • stroke / playing technique

Only three instrument classes etc.

ISSUES OF CURRENT SYSTEMS

5
slide-17
SLIDE 17

Performance not satisfying on real music Do not produce additional information for transcripts
 drum onset detection vs drum transcription

  • bars lines
  • tempo
  • meter
  • dynamics / accents
  • stroke / playing technique

Only three instrument classes etc.

ISSUES OF CURRENT SYSTEMS

5
slide-18
SLIDE 18 HH
 SD
 KD t

ADDITIONAL INFORMATION FOR TRANSCRIPTS

6
slide-19
SLIDE 19 HH
 SD
 KD t

Use beat and downbeat tracking to get:

ADDITIONAL INFORMATION FOR TRANSCRIPTS

6
slide-20
SLIDE 20 HH
 SD
 KD t 1 2 3 4 1 4 3 beats 2

Use beat and downbeat tracking to get:

  • bars lines
  • tempo
  • meter

ADDITIONAL INFORMATION FOR TRANSCRIPTS

6
slide-21
SLIDE 21 HH
 SD
 KD t 1 2 3 4 1 4 3 beats 2

Use beat and downbeat tracking to get:

  • bars lines
  • tempo
  • meter

✔

ADDITIONAL INFORMATION FOR TRANSCRIPTS

6
slide-22
SLIDE 22

IMPROVE PERFORMANCE

Three components to reach this goal:

  • 1. Leverage beat information
  • 2. Better model for drum detection
  • 3. Dataset with real music for training
7
slide-23
SLIDE 23
  • 1. LEVERAGE BEAT INFORMATION
8 HH
 SD
 KD t 1 2 3 4 1 4 3 beats 2
slide-24
SLIDE 24
  • 1. LEVERAGE BEAT INFORMATION

Beats are highly correlated with drum patterns

8 HH
 SD
 KD t 1 2 3 4 1 4 3 beats 2
slide-25
SLIDE 25
  • 1. LEVERAGE BEAT INFORMATION

Beats are highly correlated with drum patterns Assume that prior knowledge of beats is helpful for drum transcription 
 (drum hit locations / repetitive patterns)

8 HH
 SD
 KD t 1 2 3 4 1 4 3 beats 2
slide-26
SLIDE 26
  • 1. LEVERAGE BEAT INFORMATION

Beats are highly correlated with drum patterns Assume that prior knowledge of beats is helpful for drum transcription 
 (drum hit locations / repetitive patterns) Use multi-task learning for beats and drums

8 HH
 SD
 KD t 1 2 3 4 1 4 3 beats 2
slide-27
SLIDE 27

MULTI-TASK LEARNING

9 f [Hz] t [s] input
  • utput
slide-28
SLIDE 28

MULTI-TASK LEARNING

Three experiments:

9 f [Hz] t [s] input
  • utput
slide-29
SLIDE 29

MULTI-TASK LEARNING

Three experiments:

  • Training on drum targets (DT)
9 t [s] f [Hz] t [s] input
  • utput
slide-30
SLIDE 30

MULTI-TASK LEARNING

Three experiments:

  • Training on drum targets (DT)
  • Training on drum targets with annotated beats as additional input features (BF)
9 t [s] f [Hz] t [s] input
  • utput
slide-31
SLIDE 31

MULTI-TASK LEARNING

Three experiments:

  • Training on drum targets (DT)
  • Training on drum targets with annotated beats as additional input features (BF)
  • Training on drum and beat targets as multi-task problem (MT)
9 t [s] f [Hz] t [s] input
  • utput
slide-32
SLIDE 32

MULTI-TASK LEARNING

Three experiments:

  • Training on drum targets (DT)
  • Training on drum targets with annotated beats as additional input features (BF)
  • Training on drum and beat targets as multi-task problem (MT)

Expected increase in performance for BF compared to DT

9 t [s] f [Hz] t [s] input
  • utput
slide-33
SLIDE 33

MULTI-TASK LEARNING

Three experiments:

  • Training on drum targets (DT)
  • Training on drum targets with annotated beats as additional input features (BF)
  • Training on drum and beat targets as multi-task problem (MT)

Expected increase in performance for BF compared to DT Expected increase in performance for MT compared to DT

9 t [s] f [Hz] t [s] input
  • utput
slide-34
SLIDE 34
  • 2. NETWORK MODELS — BASELINE MODELS
10
slide-35
SLIDE 35

Recurrent neural networks

  • 2. NETWORK MODELS — BASELINE MODELS
10
slide-36
SLIDE 36

Recurrent neural networks

  • Recurrent connections act as memory
  • Processing of sequential data
  • 2. NETWORK MODELS — BASELINE MODELS
10 RNN train data sample
slide-37
SLIDE 37

Recurrent neural networks

  • Recurrent connections act as memory
  • Processing of sequential data
  • Work well for drum detection and beat tracking 

[Böck et al. ISMIR’16]

  • 2. NETWORK MODELS — BASELINE MODELS
10 RNN train data sample
slide-38
SLIDE 38

Recurrent neural networks

  • Recurrent connections act as memory
  • Processing of sequential data
  • Work well for drum detection and beat tracking 

[Böck et al. ISMIR’16]


RNN with label time shift (tsRNN)


state-of-the-art baseline [Vogl et al. ICASSP’17]


Bidirectional recurrent NN (BDRNN) 


[Vogl et al. ISMIR’16] [Southall et al. ISMIR’16]
  • Similar performance tsRNN
  • 2. NETWORK MODELS — BASELINE MODELS
10 RNN train data sample
slide-39
SLIDE 39
  • 2. NETWORK MODELS — NEW FOR DT
11
slide-40
SLIDE 40 CNN train data sample
  • 2. NETWORK MODELS — NEW FOR DT

Convolutional NN (CNN)

  • Convolutions capture local correlations
  • Acoustic modeling of drum sounds
11
slide-41
SLIDE 41
  • 2. NETWORK MODELS — NEW FOR DT

Convolutional NN (CNN)

  • Convolutions capture local correlations
  • Acoustic modeling of drum sounds

Convolutional BDRNN (CRNN)

  • ”best of both worlds”
  • Low-level CNN for acoustic modeling
  • Higher-level RNN for repetitive pattern modeling
11 CRNN train data sample
slide-42
SLIDE 42

NETWORK MODELS

12

Frames Context

  • Conv. Layers
  • Rec. Layers

Dense Layers BDRNN (S) 100 — — 2x50 GRU — BDRNN (L) 400 — — 3x30 GRU — CNN (S) — 9 2 x 32 3x3 filt.
 3x3 max pooling 
 2 x 64 3x3 filt.
 3x3 max pooling
 all w/ batch norm. — 2x256 CNN (L) — 25 — 2x256 CRNN (S) 100 9 2x50 GRU — CRNN (L) 400 13 3x60 GRU — tsRNN

state-of-the-art baseline [Vogl et al. ICASSP’17]
slide-43
SLIDE 43

CLASSIC DATASETS (ONLY DRUMS)

13
slide-44
SLIDE 44

IDMT-SMT-Drums [Dittmar and Gärtner 2014]

  • Solo drum tracks, recorded, synthesized, and sampled
  • 95 tracks, total: 24m, onsets: 8004 + training samples

CLASSIC DATASETS (ONLY DRUMS)

13

♫

slide-45
SLIDE 45

IDMT-SMT-Drums [Dittmar and Gärtner 2014]

  • Solo drum tracks, recorded, synthesized, and sampled
  • 95 tracks, total: 24m, onsets: 8004 + training samples

CLASSIC DATASETS (ONLY DRUMS)

13

♫

slide-46
SLIDE 46

IDMT-SMT-Drums [Dittmar and Gärtner 2014]

  • Solo drum tracks, recorded, synthesized, and sampled
  • 95 tracks, total: 24m, onsets: 8004 + training samples

ENST-Drums [Gillet and Richard 2006]

  • Recordings, three drummers on different drum kits, optional accompaniment
  • 64 tracks, total: 1h, onsets: 22391 + training samples

CLASSIC DATASETS (ONLY DRUMS)

13

♫ ♫ ♫

slide-47
SLIDE 47

IDMT-SMT-Drums [Dittmar and Gärtner 2014]

  • Solo drum tracks, recorded, synthesized, and sampled
  • 95 tracks, total: 24m, onsets: 8004 + training samples

ENST-Drums [Gillet and Richard 2006]

  • Recordings, three drummers on different drum kits, optional accompaniment
  • 64 tracks, total: 1h, onsets: 22391 + training samples

CLASSIC DATASETS (ONLY DRUMS)

13

♫ ♫ ♫

slide-48
SLIDE 48

IDMT-SMT-Drums [Dittmar and Gärtner 2014]

  • Solo drum tracks, recorded, synthesized, and sampled
  • 95 tracks, total: 24m, onsets: 8004 + training samples

ENST-Drums [Gillet and Richard 2006]

  • Recordings, three drummers on different drum kits, optional accompaniment
  • 64 tracks, total: 1h, onsets: 22391 + training samples

CLASSIC DATASETS (ONLY DRUMS)

13

♫ ♫ ♫

slide-49
SLIDE 49

DT 3-FOLD CV RESULTS ON CLASSIC DATASETS

14 F-measure [%] 60 70 80 90 100 SMT solo ENST solo ENST acc. BDRNN (S) BDRNN (L) CNN (S) CNN (L) CRNN (S) CRNN (L) tsRNN
slide-50
SLIDE 50 RBMA13-Drums [http://ifs.tuwien.ac.at/~vogl/datasets/]
  • Free music from the 2013 Red Bull Music Academy, different styles
  • 27 tracks, total: 1h 43m, onsets: 24365
  • drum, beat, and downbeat annotations

Multi-task evaluation

  • DT: Drum transcription / three fold cross-validation (same as on SMT and ENST)
  • BF: Drum transcription using annotated beats as additional input features
  • MT: Drum transcription and beat detection via multi-task learning
  • 3. NEW DATASETS (DRUMS AND BEATS)
15 NEW!

♫ ♫

slide-51
SLIDE 51 RBMA13-Drums [http://ifs.tuwien.ac.at/~vogl/datasets/]
  • Free music from the 2013 Red Bull Music Academy, different styles
  • 27 tracks, total: 1h 43m, onsets: 24365
  • drum, beat, and downbeat annotations

Multi-task evaluation

  • DT: Drum transcription / three fold cross-validation (same as on SMT and ENST)
  • BF: Drum transcription using annotated beats as additional input features
  • MT: Drum transcription and beat detection via multi-task learning
  • 3. NEW DATASETS (DRUMS AND BEATS)
15 NEW!

♫ ♫

slide-52
SLIDE 52 RBMA13-Drums [http://ifs.tuwien.ac.at/~vogl/datasets/]
  • Free music from the 2013 Red Bull Music Academy, different styles
  • 27 tracks, total: 1h 43m, onsets: 24365
  • drum, beat, and downbeat annotations

Multi-task evaluation

  • DT: Drum transcription / three fold cross-validation (same as on SMT and ENST)
  • BF: Drum transcription using annotated beats as additional input features
  • MT: Drum transcription and beat detection via multi-task learning
  • 3. NEW DATASETS (DRUMS AND BEATS)
15 NEW!

♫ ♫

slide-53
SLIDE 53 RBMA13-Drums [http://ifs.tuwien.ac.at/~vogl/datasets/]
  • Free music from the 2013 Red Bull Music Academy, different styles
  • 27 tracks, total: 1h 43m, onsets: 24365
  • drum, beat, and downbeat annotations

Multi-task evaluation

  • DT: Drum transcription / three fold cross-validation (same as on SMT and ENST)
  • BF: Drum transcription using annotated beats as additional input features
  • MT: Drum transcription and beat detection via multi-task learning
  • 3. NEW DATASETS (DRUMS AND BEATS)
15 NEW!

♫ ♫

slide-54
SLIDE 54 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) CNN (S) CNN (L) CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

RESULTS ON RBMA13

16
slide-55
SLIDE 55

RESULTS ON RBMA13: BDRNNs

17 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning
slide-56
SLIDE 56

RESULTS ON RBMA13: BDRNNs

17 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on bi-directional RNNs:

slide-57
SLIDE 57

RESULTS ON RBMA13: BDRNNs

17 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on bi-directional RNNs: BF improves for both models ✔

slide-58
SLIDE 58

RESULTS ON RBMA13: BDRNNs

17 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on bi-directional RNNs: BF improves for both models ✔ MT improves for both models ✔

slide-59
SLIDE 59

RESULTS ON RBMA13: BDRNNs

17 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on bi-directional RNNs: BF improves for both models ✔ MT improves for both models ✔ MT even better than BF for small model !

slide-60
SLIDE 60

RESULTS ON RBMA13: CNNs

18 F-measure [%] 50 55 60 65 70 CNN (S) CNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning
slide-61
SLIDE 61

RESULTS ON RBMA13: CNNs

18 F-measure [%] 50 55 60 65 70 CNN (S) CNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on CNNs:

slide-62
SLIDE 62

RESULTS ON RBMA13: CNNs

18 F-measure [%] 50 55 60 65 70 CNN (S) CNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on CNNs: BF inconsistent

slide-63
SLIDE 63

RESULTS ON RBMA13: CNNs

18 F-measure [%] 50 55 60 65 70 CNN (S) CNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on CNNs: BF inconsistent MT declines for both models

slide-64
SLIDE 64

RESULTS ON RBMA13: CRNNs

19 F-measure [%] 50 55 60 65 70 CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning
slide-65
SLIDE 65

RESULTS ON RBMA13: CRNNs

19 F-measure [%] 50 55 60 65 70 CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on CRNNs:

slide-66
SLIDE 66

RESULTS ON RBMA13: CRNNs

19 F-measure [%] 50 55 60 65 70 CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on CRNNs: BF improves for both models ✔

slide-67
SLIDE 67

RESULTS ON RBMA13: CRNNs

19 F-measure [%] 50 55 60 65 70 CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on CRNNs: BF improves for both models ✔ MT improves for small models ✔

slide-68
SLIDE 68

RESULTS ON RBMA13: CRNNs

19 F-measure [%] 50 55 60 65 70 CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on CRNNs: BF improves for both models ✔ MT improves for small models ✔ MT even better than BF for small model !

slide-69
SLIDE 69

RESULTS ON RBMA13: CRNNs

19 F-measure [%] 50 55 60 65 70 CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

Impact on CRNNs: BF improves for both models ✔ MT improves for small models ✔ MT even better than BF for small model ! MT equal for large model ?

slide-70
SLIDE 70 20 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

RESULTS FOR RECURRENT ARCHITECTURES

slide-71
SLIDE 71 20 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

RESULTS FOR RECURRENT ARCHITECTURES

slide-72
SLIDE 72 20 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

RESULTS FOR RECURRENT ARCHITECTURES

slide-73
SLIDE 73 20 F-measure [%] 50 55 60 65 70 BDRNN (S) BDRNN (L) CRNN (S) CRNN (L) DT … Drum transcription (3-fold CV) BF … Drum transcription using annotated beats as additional input features MT … Drum transcription and beat detection via multi-task learning

RESULTS FOR RECURRENT ARCHITECTURES

No improvement because of 
 beat tracking results?
slide-74
SLIDE 74

CONCLUSIONS

21
slide-75
SLIDE 75

CONCLUSIONS

Use beats and downbeats to get meta information for transcripts

21
slide-76
SLIDE 76

CONCLUSIONS

Use beats and downbeats to get meta information for transcripts Multi-task learning for drums and beats can be beneficial for recurrent architectures

21
slide-77
SLIDE 77

CONCLUSIONS

Use beats and downbeats to get meta information for transcripts Multi-task learning for drums and beats can be beneficial for recurrent architectures CRNNs can outperform RNNs

21
slide-78
SLIDE 78

CONCLUSIONS

Use beats and downbeats to get meta information for transcripts Multi-task learning for drums and beats can be beneficial for recurrent architectures CRNNs can outperform RNNs CRNN best overall results @ MIREX’17 drum transcription


MIREX system: http://ifs.tuwien.ac.at/~vogl/models/mirex-17.zip
 madmom: https://github.com/CPJKU/madmom 21
slide-79
SLIDE 79

CONCLUSIONS

Use beats and downbeats to get meta information for transcripts Multi-task learning for drums and beats can be beneficial for recurrent architectures CRNNs can outperform RNNs CRNN best overall results @ MIREX’17 drum transcription


MIREX system: http://ifs.tuwien.ac.at/~vogl/models/mirex-17.zip
 madmom: https://github.com/CPJKU/madmom

New dataset with free music featuring beat, and drum annotations


http://ifs.tuwien.ac.at/~vogl/datasets/ 21
slide-80
SLIDE 80

CONCLUSIONS

Use beats and downbeats to get meta information for transcripts Multi-task learning for drums and beats can be beneficial for recurrent architectures CRNNs can outperform RNNs CRNN best overall results @ MIREX’17 drum transcription


MIREX system: http://ifs.tuwien.ac.at/~vogl/models/mirex-17.zip
 madmom: https://github.com/CPJKU/madmom

New dataset with free music featuring beat, and drum annotations


http://ifs.tuwien.ac.at/~vogl/datasets/ 21