Fooling Singing Voice Detection CNNs with Hand-Drawn Examples

For the presentation of my ISMIR 2016 paper titled "Learning to Pinpoint Singing Voice from Weakly-Labeled Examples" [Paper, BibTeX], I put together a small web demo showing how a state-of-the-art singing voice detector seems to only rely on detecting sloped or wiggly lines in the spectrogram.

The web demo allows you to draw into a mel spectrogram of a piano recording, compute singing voice predictions with a CNN, and also listen to what you drew:

It is currently available online at https://jobim.ofai.at/singinghorse (the first computation might take quite a while, in case it has to reload Theano and recompile the function, others will take less than a second).

The code and model in this repository allows you to run the demo locally on your machine instead, or take a look at how it's done. Feel free to copy out bits and pieces, but make sure to obey and include the license.

For the baseline experiments of my ISMIR 2015 paper, see the ismir2015 repository, and for experiments on training a network to not be irritated by wiggly lines, see its unhorse branch.

Preliminaries

Running the code locally requires the following software:

Python 2.7+ or 3.4+
Python packages: numpy, scipy, matplotlib, h5py
ffmpeg or avconv

For better performance, the following Python packages are recommended:

Theano, Lasagne (for much faster network predictions)
pyfftw (for much faster spectrogram computation)
bjoern (for a more stable web server)

For Theano and Lasagne, you may need the bleeding-edge versions from github. In short, they can be installed with:

pip install --upgrade --no-deps https://github.com/Theano/Theano/archive/master.zip
pip install --upgrade --no-deps https://github.com/Lasagne/Lasagne/archive/master.zip

(Add --user to install in your home directory, or sudo to install globally.) For more complete installation instructions including GPU setup, please refer to the From Zero to Lasagne guides.

On Ubuntu, pyfftw can be installed with the following two commands:

sudo apt-get install libfftw3-dev
pip install pyfftw

Finally, bjoern is available on PyPI:

pip install bjoern

(Again, optionally add --user or sudo.)

Running

Clone the repository somewhere:

git clone https://github.com/f0k/singing_horse.git

If you do not have git available, download the code from https://github.com/f0k/singing_horse/archive/master.zip and extract it.

Run the built-in web server with:

python serve.py

You should now be able to access it at http://localhost:9090.

Alternatively, under Apache, activate mod_wsgi and add a WSGIScriptAlias directive mapping some URL to serve.py.

About...

... the implementation

The spectrogram <img> is overlaid with a same-sized <canvas> element that is semi-transparent. Mouse movements are translated to line drawings with some minimal JavaScript. Clicking the "Compute" button uploads the line drawing as a PNG data URI to the web server, via a POST request to /render. In serve.py, this URI is decoded to binary, then decoded to a numpy array with the Python Imaging Library (PIL). It is combined with the (precomputed) spectrogram of the test audio snippet, and passed to the CNN in predict.py. The resulting prediction curve is rendered to an in-memory PNG with matplotlib in plot.py. Furthermore, the combined spectrogram is resynthesized to audio samples in spects.py, and then encoded to an in-memory MP3 file with ffmpeg. Finally, the MP3 file and PNG are encoded as data URIs and sent back to the browser, which updates the <img> and <audio> sources with the data URIs. (The data is small enough for data URIs, and sending it directly to the client in this way avoids having to store and manage temporary image and audio files on the server.)

... the code

The code was thrown together for the demonstration in my ISMIR 2016 presentation, and is not fully-polished and fully-documented. I am publishing it because several people asked how this was done, and I hope it still serves to explain this. Feel free to open an issue here or send me an email if you have any questions. If you reuse parts of the code for your purposes, please make sure to include the license.