Need help with Speech-enhancement?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

208 Stars 55 Forks MIT License 27 Commits 7 Opened issues


Deep learning for audio denoising

Services available


Need anything else?

Contributors list


Build Status

Vincent Belz : [email protected]

Published in towards data science : Speech-enhancement with Deep learning


This project aims at building a speech enhancement system to attenuate environmental noise.

Spectrogram denoising

Audios have many different ways to be represented, going from raw time series to time-frequency decompositions. The choice of the representation is crucial for the performance of your system. Among time-frequency decompositions, Spectrograms have been proved to be a useful representation for audio processing. They consist in 2D images representing sequences of Short Time Fourier Transform (STFT) with time and frequency as axes, and brightness representing the strength of a frequency component at each time frame. In such they appear a natural domain to apply the CNNS architectures for images directly to sound. Between magnitude and phase spectrograms, magnitude spectrograms contain most the structure of the signal. Phase spectrograms appear to show only little temporal and spectral regularities.

In this project, I will use magnitude spectrograms as a representation of sound (cf image below) in order to predict the noise model to be subtracted to a noisy voice spectrogram.

sound representation

The project is decomposed in three modes:

data creation

Prepare the data

To create the datasets for training, I gathered english speech clean voices and environmental noises from different sources.

The clean voices were mainly gathered from LibriSpeech: an ASR corpus based on public domain audio books. I used as well some datas from SiSec. The environmental noises were gathered from ESC-50 dataset or

For this project, I focused on 10 classes of environmental noise: tic clock, foot steps, bells, handsaw, alarm, fireworks, insects, brushing teeth, vaccum cleaner and snoring. These classes are illustrated in the image below (I created this image using pictures from

classes of environmental noise used

To create the datasets for training/validation/testing, audios were sampled at 8kHz and I extracted windows slighly above 1 second. I performed some data augmentation for the environmental noises (taking the windows at different times creates different noise windows). Noises have been blended to clean voices with a randomization of the noise level (between 20% and 80%). At the end, training data consisted of 10h of noisy voice & clean voice, and validation data of 1h of sound.

To prepare the data, I recommend to create data/Train and data/Test folders in a location separate from your code folder. Then create the following structure as in the image below:

data folder structure

You would modify the

, and
paths name accordingly into the
file that takes the default parameters for the program.

Place your noise audio files into

directory and your clean voice files into

Specify how many frames you want to create as

(or pass it as argument from the terminal) I let nb_samples=50 by default for the demo but for production I would recommend having 40 000 or more.

Then run

python --mode='data_creation'
. This will randomly blend some clean voices from
with some noises from
and save the spectrograms of noisy voices, noises and clean voices to disk as well as complex phases, time series and sounds (for QC or to test other networks). It takes the inputs parameters defined in
. Parameters for STFT, frame length, hop_length can be modified in
(or pass it as arguments from the terminal), but with the default parameters each window will be converted into spectrogram matrix of size 128 x 128.

Datasets to be used for training will be magnitude spectrograms of noisy voices and magnitude spectrograms of clean voices.


The model used for the training is a U-Net, a Deep Convolutional Autoencoder with symmetric skip connections. U-Net was initially developed for Bio Medical Image Segmentation. Here the U-Net has been adapted to denoise spectrograms.

As input to the network, the magnitude spectrograms of the noisy voices. As output the Noise to model (noisy voice magnitude spectrogram - clean voice magnitude spectrogram). Both input and output matrix are scaled with a global scaling to be mapped into a distribution between -1 and 1.

Unet training

Many configurations have been tested during the training. For the preferred configuration the encoder is made of 10 convolutional layers (with LeakyReLU, maxpooling and dropout). The decoder is a symmetric expanding path with skip connections. The last activation layer is a hyperbolic tangent (tanh) to have an output distribution between -1 and 1. For training from scratch the initial random weights where set with He normal initializer.

Model is compiled with Adam optimizer and the loss function used is the Huber loss as a compromise between the L1 and L2 loss.

Training on a modern GPU takes a couple of hours.

If you have a GPU for deep learning computation in your local computer, you can train with:

python --mode="training"
. It takes as inputs parameters defined in
. By default it will train from scratch (you can change this by turning
to false). You can start training from pre-trained weights specified in
. I let available
with weights from my training in
. The number of epochs and the batch size for training are specified by
. Best weights are automatically saved during training as
. You can call fit_generator to only load part of the data to disk at training time.

Personally, I used the free GPU available at Google colab for my training. I let a notebook example at

. If you have a large available space on your drive, you can load all your training data to your drive and load part of it at training time with the fit_generator option of tensorflow.keras. Personally I had limited space available on my Google drive so I pre-prepared in advanced batches of 5Gb to be loaded to drive for training. Weights were regularly saved and reload for next training.

At the end, I obtained a training loss of 0.002129 and a validation loss of 0.002406. Below a loss graph made in one of the trainings.

loss training


For prediction, the noisy voice audios are converted into numpy time series of windows slightly above 1 second. Each time serie is converted into a magnitude spectrogram and a phase spectrogram via STFT transforms. Noisy voice spectrograms are passed into the U-Net network that will predict the noise model for each window (cf graph below). Prediction time for one window once converted to magnitude spectrogram is around 80 ms using classical CPU.

flow prediction part 1

Then the model is subtracted from the noisy voice spectrogram (here I apply a direct subtraction as it was sufficient for my task, we could imagine to train a second network to adapt the noise model, or applying a matching filter such as performed in signal processing). The "denoised" magnitude spectrogram is combined with the initial phase as input for the inverse Short Time Fourier Transform (ISTFT). Our denoised time serie can be then converted to audio (cf graph below).

flow prediction part 2

Let's have a look at the performance on validation datas!

Below I display some results from validation examples for Alarm/Insects/Vaccum cleaner/Bells noise. For each of them I display the initial noisy voice spectrogram, the denoised spectrogram predicted by the network, and the true clean voice spectrogram. We can see that the network is well able to generalize the noise modelling, and produce a slightly smoothed version of the voice spectrogram, very close to the true clean voice spectrogram.

More examples of spectrogram denoising on validation data are displayed in the initial gif on top of the repository.

validation examples

Let's hear the results converted back to sounds:

Audios for Alarm example:

Input example alarm

Predicted output example alarm

True output example alarm

Audios for Insects example:

Input example insects

Predicted output example insects

True output example insects

Audios for Vaccum cleaner example:

Input example vaccum cleaner

Predicted output example vaccum cleaner

True output example vaccum cleaner

Audios for Bells example:

Input example bells

Predicted output example bells

True output example bells

Below I show the corresponding displays converting back to time series:

validation examples timeserie

You can have a look at these displays/audios in the jupyter notebook

that I provide in the

Below, I show the corresponding gif of the spectrogram denoising gif (top of the repository) in the time serie domain.

Timeserie denoising

As an extreme testing, I applied to some voices blended with many noises at a high level. The network appeared to work surprisingly well for the denoising. The total time to denoise a 5 seconds audio was around 4 seconds (using classical CPU).

Below some examples:

Example 1:

Input example test 1

Predicted output example test 1

Example 2:

Input example test 2

Predicted output example test 2

How to use?

- Clone this repository
- pip install -r requirements.txt
- python OPTIONS

  • Modes of the program (Possible OPTIONS):

--mode: default='prediction', type=str, choices=['data_creation', 'training', 'prediction']

Have a look at possible arguments for each option in


Jansson, Andreas, Eric J. Humphrey, Nicola Montecchio, Rachel M. Bittner, Aparna Kumar and Tillman Weyde.Singing Voice Separation with Deep U-Net Convolutional Networks. ISMIR (2017).


Grais, Emad M. and Plumbley, Mark D., Single Channel Audio Source Separation using Convolutional Denoising Autoencoders (2017).


Ronneberger O., Fischer P., Brox T. (2015) U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N., Hornegger J., Wells W., Frangi A. (eds) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. MICCAI 2015. Lecture Notes in Computer Science, vol 9351. Springer, Cham


K. J. Piczak. ESC: Dataset for Environmental Sound Classification. Proceedings of the 23rd Annual ACM Conference on Multimedia, Brisbane, Australia, 2015.




We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.