∂SMMC experiments with singing voice

Juan Alonso and I have published the first part of differentiable DSP experiments with singing voice. With these experiments, you can voice-clone a speaker to sing your favourite song with about 15 minutes of recorded speech.

The code is available at https://github.com/juanalonso/DDSP-singing-experiments with links to the zero-configuration notebooks, and sound examples are at
https://juanalonso.github.io/DDSP-singing-experiments/ .

Preprint is at : http://arxiv.org/abs/2103.07197 

You can see the twitter thread in Spanish here: https://twitter.com/kokuma/status/1371799702413381632

and the English mini thread here:

Happy experimenting.

Towards Differentiable Sound and Music Computing

Deep Learning (DL) had a big impact in digitalization. When big data, computational power, good objectives and metrics are available for certain tasks, the DL structures learn quite accurately without explicit programming. However, in real-world of sound, music, and movement computing (SMMC), these ideal components almost never come together. Moreover, as the resources grow, specialize and become varied day by day, it becomes a challenge to keep up, develop, evaluate, deploy, and maintain useful

The DL structures can alternatively be interpreted as computational models, such as Turing machines, where CPU-like neural controllers access external memory for learning real-world SMMC tasks. This representation should be familiar to SMMC researchers who work on embedded platforms such as digital signal processing (DSP) chips. Then, say, a 1-D convolutional neural network becomes a non-recursive filter, and a recurrent neural network becomes a recursive filter. The benefits of this approach are: 1)
Differentiability, 2) domain specific components, and familiarity 3) Less data, more domain knowledge (inductive biases) 4) Less training, and 5) continuous adaptation to the input data.

Building neural network structures can thus be approached in a similar way to building traditional signal processing chains, making it possible to use gradient descent to set the parameters of these chains. These DSP chains are just one example of acyclic graph structures, there are many others in SMMC. Examples include data-flow programming languages, block-compilers, Open AL, motion capture data structures that have skeleton embeddings (bvf, fbx, etc), and Open Scene Graph for AR/VR.

With the new advances in differentiable programming (∂P), it becomes feasible and timely to choose the best domain-specific code for each chain by gradient descent (Software 2.0). Within this context, I aim to *understand* how ∂P can generate rich, interactive, synthetic audio-visual media that can be applied to sound, music, and movement computing (∂SMMC).

Epiphany #24 (1981/82)

I am programming my first computer, the Commodore 64. My first sound program is something like this. I command to the SID-chip with direct memory access, via PEEK and POKE commands.

The memory locations (MEM) used for SID music synthesis start at 54272 ($D400) in the Commodore 64. The memory locations 54272 to 54296 inclusive are the POKE locations you need to remember when you’re using the 6581 (SID) chip register map.

When I run the program, there are weird noises. I realize I have typed a memory location wrong, I am reading & writing to a random region.

I think: could I use this to get a metallic tinge-like sound like Kraftwerk? I discovered them at the same year on the radio.

start tok64 page185.prg
  5 s=54272
  10 for l=s to s+24:poke l,0:next    :rem clear sound chip
  20 poke s+5,9:poke s+6,0
  30 poke s+24,15              :rem set volume to maximum
  40 read hf,lf,dr
  50 if hf<0 then end
  60 poke s+1, hf :poke s, lf
  70 poke s+4, 33
  80 for t = 1 to dr :next
  90 poke s+4,32 :for t=1 to 50:next
  100 goto 40
  110 data 25,177,250,28,214,250
  120 data 25,177,250,25,177,250
  130 data 25,177,125,28,214,125
  140 data 32,94,750,25,177,250
  150 data 28,214,250,19,63,250
  160 data 19,63,250,19,63,250
  170 data 21,154,63,24,63,63
  180 data 25,177,250,24,63,125
  190 data 19,63,250,-1,-1,-1
stop tok64

This is one of the reasons early / assembler programmers like me get why Neural Turning Machines [1], Differentiable Neural Computers [2] work as they do. More to come on this topic.


[1] Graves, A., Wayne, G., Danihelka, I. (2014). Neural Turing Machines https://arxiv.org/abs/1410.5401

[2]  Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwińska, A., Colmenarejo, S., Grefenstette, E., Ramalho, T., Agapiou, J., Badia, A., Hermann, K., Zwols, Y., Ostrovski, G., Cain, A., King, H., Summerfield, C., Blunsom, P., Kavukcuoglu, K., Hassabis, D. (2016). Hybrid computing using a neural network with dynamic external memory Nature 538(7626), 471-476. https://dx.doi.org/10.1038/nature20101