SinGlow: Generative Flow for SVS tasks in Tensorflow 2

python 3 tensorflow 2

See more in the paper: SinGlow: Singing Voice Synthesis with Glow —- Help Virtual Singers More Human-like

SinGlow is a part of my Singing voice synthesis system. It can extract features of sound, particularly songs and musics. Then we can use these features (or perfect encoding) for feature migrating tasks. For example migrate features of real singers’ song to those virtual singers’ songs.

This project is developed above the project GLOW-tf2 under MIT licence, and the following words are from its developers.

My implementation of GLOW from the paper in Tensorflow 2. GLOW is an interesting generative model as it uses invertible neural network to transform images to normal distribution and vice versa. Additionally, it is strongly based on RealNVP, so knowing it would be helpful to understand GLOW’s contribution.


Singing voice synthesis (SVS) is a task using the computer to generate songs with lyrics. So far, researchers are focusing on tunning the pre-recorded sound pieces according to rigid rules. For example, in Vocaloid, one of the commercial SVS systems, there are 8 principal parameters modifiable by song creators. The system uses these parameters to synthesise sound pieces pre-recorded from professional voice actors. We notice a common difference between computer-generated songs and real singers’ songs. This difference can be addressed to help the generated ones become more like the real-singer ones.

In this paper, we propose SinGlow, as a solution to minimise this difference. SinGlow is one of the Normalising Flow that directly uses the calculated Negative Log-Likelihood value to optimise the trainable parameters. This feature gives SinGlow the ability to perfectly encode inputs into feature vectors, which allows us to manipulate the feature space to minimise the difference we discussed before. To our best knowledge, we are the first to propose an application of Normalising Flow in SVS fields.

In our experiments, SinGlow shows the ability to encode sound and make the input virtual-singer songs more human-like.


│ //need modification, replace data dirs with yours
│ //model configurations are located here
│ //construct tfrecord dataset from wav or mp3 data, and load it
│ //Glow Model / SinGlow Model
│ //training pipeline
│ //originate from Glow-OpenAI
│ //originate from Tensorflow
│       weights.h5 //the model weights file
├───runs //outputs and rerecords
├───logs //tensorboard logdir
├───design //model architecture information
        run.ipynb //dataset construction and applying model
        experiment.ipynb //evaluate model //some user guide


pip3 install -r requirements.txt


After every epoch, the network’s weights will be stored in the checkpoints directory defined in

There are also some sampling of the network (image generation mode) that are going to be stored in results directory. Additionally, TensorBoard is used to track z’s mean and variance, as well as the negative log-likelihood.

In optimal state, z should have zero mean and one variance. Additionally, the TensorBoard stores sampling with temperature of 0.7.

python3 [-h] [--dataset [DATASET]] [--k_glow [K_GLOW]] [--l_glow [L_GLOW]]
       [--img_size [IMG_SIZE]] [--channel_size [CHANNEL_SIZE]]

optional arguments:
  -h, --help            show this help message and exit
  --dataset [DATASET]   The dataset to train on ("mnist", "cifar10", "cifar100")
  --k_glow [K_GLOW]     The amount of blocks per layer
  --l_glow [L_GLOW]     The amount of layers
  --img_size [IMG_SIZE] The width and height of the input images
  --channel_size [CHANNEL_SIZE]
                        The channel size of the input images


To contribute to the project, these steps can be followed. Anyone that contributes will surely be recognized and mentioned here!

Contributions to the project are made using the “Fork & Pull” model. The typical steps would be:

  1. create an account on github
  2. fork this repository
  3. make a local clone
  4. make changes on the local copy
  5. commit changes git commit -m "my message"
  6. push to your GitHub account: git push origin
  7. create a Pull Request (PR) from your GitHub fork
    (go to your fork’s webpage and click on “Pull Request.”
    You can then add a message to describe your proposal.)


This open-source project is licensed under MIT License.


TODO the reference information