Generating Music with GANs: An Overview and Case Studies by Hao-Wen Dong and Yi-Hsuan Yang.
Waveform-based music processing with deep learning by Sander Dieleman, Jordi Pons and Jongpil Lee. I have blogged a bunch of Jordi’s work here under source separation. Sander’s presentation had some interesting framings about
- mode-seeking versus mode-covering approximations to probablility distributions.
- sparse versus densley conditioned conditional signals
Papers that are useful for my own interests, that is; this is not necessaarily an indictment of any papers I do not mention.
Or… See the ISMIR paper explorer.
- Obviously I like my paper (MacKinlay and Botev 2019) and think it is the best and most eloquently explained.
- Keunwoo Choi’s Drummernet (K. Choi and Cho 2019) looks like a cunning hack to transcribe drums from audio, by learning to play a drum synthesizer.
- (J. Choi et al. 2019) claims to solve a lot of the notorious problems with noisy labelling in music with a zero short learning model.
- Stefan Lattner’s Drumnet (Lattner and Grachten 2019) is a remarkably simple model for rhythm generation.
- Magdalena Fuentes et al, on detecting microtime in Afro Latin rhythms was super fun (Fuentes et al. 2019).
- Work by Ashis Pati et al is nice. Learning to Traverse Latent Spaces for Musical Score Inpainting (Pati, Lerch, and Hadjeres 2019).
- Generating Structured Drum Pattern Using Variational Autoencoder and Self-similarity Matrix (Wei, Wu, and Su 2019) I hope to track these folks down but we are presenting our research at the same time. But this covariance structure appeals to me.
- Supervised symbolic music style translation using synthetic data (Ondrˇej Cífka and Richard 2019) is kind of an automated Señor Coconut.
- Spleeter (Hennequin et al. 2019) from Deezer labs is one deep learning approach
- Open Unmix (Stöter et al. 2019) from Sony CSL labs is another deep learning apprach
- UNMIXER (Smith, Kawasaki, and Goto 2019) a web UI for a cute hand-rolled matrix factorisation method
All bloggged under source separation.
A lot of the authors would like to impose a certain factorisation, or “near”-factorisation, over a latent space into humanly interpretable dimensions. So they would like to disentangle, say, timbre from pitch from loudness, or similar. I would like to return to this problem; It looks fun.
- Coupled Recurrent Models for Polyphonic Music Composition (Thickstun et al. 2019). It is phrased as a neural network problem, but their central question is, to my mind: What graphical model structure best approximates polyphonic scores?
- Hanoi Hantrakul presenting Fast and Flexible Neural Audio Synthesis. The oral presentaiton turned out to be an advertisement for the successor project, Differentiable DSP.
- Yin-Jyun Luo et al have done something interesting in Learning Disentangled Representations of Timbre and Pitch for Musical Instrument Sounds Using Gaussian Mixture Variational Autoencoders Check out the demo page. (Luo, Agres, and Herremans 2019)
- Learning Complex Basis Functions for Invariant Representations of Audio (Lattner, Dorﬂer, and Arzt 2019); Here it is about finding basis function which preserve a priori symmetries. Appended to the sparse coding page.
- Deep Music Analogy Via Latent Representation Disentanglement (Yang et al. 2019)
- mirdata, (Bittner et al. 2019)
- The AcousticBrainz Genre Dataset (Bogdanov et al. 2019)
- Harmonix set (Nieto et al. 2019)
- AIST Dance DB Dance videos! (Tsuchida et al. 2019)
All blogged under audio corpora.
- https://acids-ircam.github.io/flow_synthesizer/ (Esling et al. 2019)
The So Strangely music science podcast.