AudioOnsetDetection

The test dataset from:

Molina E , Tardon L J , Barbancho A M , et al. SiPTH: Singing Transcription Based on Hysteresis Defined on the Pitch-Time Curve[J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2015, 23(2):252-263.
The onset detection only check the onset time in the ground-truth files

The train dataset from:

Salamon J , Joan Serrà, Emilia Gómez. Tonal representations for music retrieval: From version identification to query-by-humming[J]. International Journal of Multimedia Information Retrieval, 2013.
http://mtg.upf.edu/download/datasets/MTG-QBH

Tips of onset detection.exe

cmd> onset_detection.exe *.wav
output *csv file with time
0.011
1.222
3.222

Evaluation result

raw data of baseline and verification

Approach	precision	recall	f1-measure
onset function(CNN)	0.646	0.863	0.732
onset function(RNN)	0.711	0.686	0.671
baseline	0.414	0.922	0.564
our CNN	0.916	0.836	0.872
our RNN	0.915	0.850	0.880
our LSTM	0.926	0.850	0.885

CNN\RNN onset function is the state-of-art approaches that construct a function to represent onset versus non-onset. However, the recall decrease while the precision increses that would have less potential to detect more onset, i.e., the we could not get an acceptable F1-measure. Our aprraches have more accuray with less change of recall, then get more F1-measure.

onset function(CNN):

[1]Musical Onset Detection with Convolutional Neural Networks” Jan Schlüter and Sebastian Böck. Proceedings of the 6th International Workshop on Machine Learning and Music, 2013.

onset function(RNN):

[1] “Universal Onset Detection with bidirectional Long Short-Term Memory Neural Networks” Florian Eyben, Sebastian Böck, Björn Schuller and Alex Graves. Proceedings of the 11th International Society for Music Information Retrieval Conference (ISMIR), 2010.
[2] “Online Real-time Onset Detection with Recurrent Neural Networks” Sebastian Böck, Andreas Arzt, Florian Krebs and Markus Schedl. Proceedings of the 15th International Conference on Digital Audio Effects (DAFx), 2012.

our approach that add a verification step to the onset detection:

baseline:we modify the tradidtional pipelines of peak-picking with sliding window and methods of post-processing with deep neural network.
our CNN,our RNN,our LSTM: we train the classifiction model of neural network after our baseline approach.

---

Training sets and annotations, as well as network training source code, will be released in the near furture.