We present a new machine learning technique for generating music and audio signals. The focus of this work is to develop new techniques parallel to what has been proposed for artistic style transfer for images by Gatys et al. . We present two cases of modifying an audio signal to generate new sounds. A feature of our method is that a single architecture can generate these different audio-style-transfer types using the same set of parameters which otherwise require complex hand-tuned diverse signal processing pipelines. Finally, we propose and investigate generation of spectrograms from noise by satisfying an optimization criterion derived from features derived from filter-activations of a convolutional neural net. The potential flexibility of this sound-generating approach is discussed.
There has been recent work applying architectures in computer vision for acoustic scene analysis. In particular,  uses standard architectures such as AlexNet, VGG-Net, and ResNet for sound understanding. The performance gains from the vision models are translated to the audio domain as well. The work in [2, 3] used a mel-filter-bank input representation, while we use Short-Time Fourier Transform (STFT) log-magnitude instead. We desire a high-resolution audio representation from which perfect reconstruction is possible via, e.g., Griffin-Lim Reconstruction . All experiments in this work use an audio spectrogram representation having duration 2.57s, frame-size 30ms, frame-step 10ms, FFT-size 512, and audio sampling rate of 16kHz.
The core of the success of neural style transfer for vision is to optimize the input signal, starting with random noise, to take on the features of interest derived from activations at different layers after the passing through a convolutional net based classifier which was trained on the content of the input image. We follow a similar approach, with some modifications for audio signals. First, we train a standard AlexNet  architecture, but have a smaller receptive size ofinstead of the larger receptive fields used in the original work. This is to retain the audio resolution, both along time and frequency, as larger receptive fields would yield poor localization in the audio reconstruction, which results in audible artifacts. We also add additional loss terms in order to match the averaged timbral and energy envelope. All applications here correspond to timbre transfer of musical instruments having no explicit knowledge of features such as pitch, note onset time, type of instrument, and so on. The AlexNet was trained on audio spectograms to distinguish classes of musical instrument sounds ( from AudioSet), with convolutions and pooling, having a total of layers with objective function minimizing the cross-entropy loss using the Adam optimizer .
We focus on two experiments: (1) imposing the style of a tuning fork on a harp, resulting in bandwidth compression down to the fundamental, and (2) transferring the style of a violin note to a singing voice, resulting in bandwidth expansion. Thus, we have a new form of cross-synthesis imposing the style of one instrument on the content of another, with applications similar to . We explored various hyper-parameters and single/multiple layers from which we extract these features for optimization. The goal is to have a single parameter setting that can perform all of these tasks, without having to explicitly develop hand-crafted rules. Traditionally there have been distinct signal processing based approaches to do such tasks. Subplots in Figs. 1-2 a)-d) are log-magnitude spectrograms with the y-axis 0-8kHz and x-axis 0-2.57s. Note in Fig. 2. how this approach not only changes the timbre, but also increases the bandwidth of the signal, as seen in the strength of the higher harmonics. The objective equation below drives the reconstructed spectrogram from random noise to be the spectra that minimizes the sum of weighted loss terms denoting the content loss (the Euclidean norm of the difference between the current activation filters and those of the content spectrogram), the style loss (which is the normalized Euclidean norm between the Gram matrix of filter activations of selected convolutional layers similar to  between and ), and and which measure deviation in the temporal and frequency energy envelopes respectively from the style audio. We found that matching the weighted energy contour and frequency energy contour (timbral envelope), namely and
, averaged over time in our loss function, helped in achieving improved quality. The energy term in the loss function is required because the Gram matrix does not incorporate temporal dynamics of the target audio style, and would generally follow that of the content if not included.
4 Conclusion and Future Work
We have proposed a novel way to synthesize audio by treating it as a
style-transfer problem, starting from a random-noise input signal and
iteratively using back-propagation to optimize the sound to conform to
filter-outputs from a pre-trained neural architecture. The two
examples were intended to explore and illustrate the nature of the
style transfer for spectrograms, and more musical examples are
subjects of ongoing work. The flexibility of this approach, and the
promising results to date indicate interesting future sound
cross-synthesis methods. We believe this work can be extended to
many new audio synthesis/modification techniques based on new
loss-term formulations for the problem of interest, and are excited to see and hear what lies
111 Acknowledgments: The authors would like to thank Andrew Ng’s group and the Stanford Artificial Intelligence Laboratory for the use of their computing resources. Prateek Verma would like to thank Ziang Xie for discussion about challenges in the problem, and Alexandre Alahi for style transfer work in computer vision.
Acknowledgments: The authors would like to thank Andrew Ng’s group and the Stanford Artificial Intelligence Laboratory for the use of their computing resources. Prateek Verma would like to thank Ziang Xie for discussion about challenges in the problem, and Alexandre Alahi for style transfer work in computer vision.
Gatys, Leon A., Alexander S. Ecker, and Matthias Bethge. "A neural algorithm of artistic style." arXiv preprint arXiv:1508.06576(2015).
Recommending music on Spotify with deep learning – Sander Dieleman benanne.github.io/2014/08/05/spotify-cnns.html
Hershey, Shawn, et al. "CNN architectures for large-scale audio classification." Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE International Conference on. IEEE, 2017.
Kingma, Diederik, and Jimmy Ba. "Adam: A method for stochastic optimization." arXiv preprint arXiv:1412.6980 (2014).
Verma, Prateek, and Preeti Rao. "Real-time melodic accompaniment system for Indian music using TMS320C6713." VLSI Design (VLSID), 2012 25th International Conference on. IEEE, 2012.
Griffin, Daniel, and Jae Lim. "Signal estimation from modified short-time Fourier transform." IEEE Transactions on Acoustics, Speech, and Signal Processing 32.2 (1984): 236-243.