Can anyone please share a matlab code for reverberation or feature perturbation for data augmentation purpose. I need it for Speaker Recognition task.
Hello Mr., Kawthar Yasmine Zergat,
I do not have a specific Matlab code, but the following one is from python. The code is really easy to understand. the X_input and y_input, are the features and label data-sets, respectively. Just make some changes and you will have a nice code.
def data_augmentation(X_input, y_input, t_degree=4):
# Function to augment the data
xs, xn = [], []
# The positive data or "1" are going to be augmented n times
for i in range(t_degree):
mask = y_input>0
x1 = X_input[mask].copy()
ids = np.arange(x1.shape[0])
for c in range(x1.shape[1]):
np.random.shuffle(ids)
x1[:, c] = x1[ids][:, c]
xs.append(x1)
# The data targeted as 0, are not going to augmented
for i in range(t_degree):
mask = y_input==0
x1 = X_input[mask].copy()
ids = np.arange(x1.shape[0])
for c in range(x1.shape[1]):
np.random.shuffle(ids)
x1[:, c]=x1[ids][:, c]
xn.append(x1)
xs = np.vstack(xs)
xn = np.vstack(xn)
ys = np.ones(xs.shape[0])
yn = np.zeros(xn.shape[0])
X = np.vstack([X_input,xs, xn])
y = np.concatenate([y_input,ys, yn])
return X,y
Best regards,
get from here
https://github.com/codebox/image_augmentor
https://towardsdatascience.com/image-augmentation-examples-in-python-d552c26f2873
https://www.mathworks.com/help/deeplearning/ref/imagedataaugmenter.html
Dear kumar, thank you for your answer. I dont know if I can use the same process as image augmentation.
Hi Kawthar,
I am afraid to say that you may not get any publicly available code in MATLAB for speech data augmentation.
You can some examples used for KALDI here (Stage 2) https://github.com/kaldi-asr/kaldi/blob/master/egs/voxceleb/v2/run.sh
It's not hard to replicate them (line no 72 for reveraberation) in MATLAB.
Hope this helps.
Regards
Sahid
This project seems to fit your needs :
https://code.soundsoftware.ac.uk/projects/audio-degradation-toolbox
-
Otherwise, there are many python implementations on Github for audio degradation such as :
https://github.com/EliosMolina/audio_degrader
-
Another solution would be to download / compile kaldi and use the wav-reverberate binary directly :
e.g. (from the kaldi documentation) :
wav-reverberate --duration=20.25 --impulse-response=rir.wav --additive-signals='noise1.wav,noise2.wav' --snrs='20.0,15.0' --start-times='0,17.8' input.wav output.wav
(you can use musan from room impulse responses : http://www.openslr.org/28/)
Salma Lahmar Data augmentation can be applied to any kind of signal (or data). The goal is to generalize to unseen data when a limited variety of conditions is observed in the dataset. This principle is applicable to images (rotation, zoom, flip, etc), speech (adding noises, reverberation, changing the speed/pitch, etc), videos (chunking, subsampling, etc) and can be performed in a image-like domain where the data is 2-dimensional (spectrogram for speech / frame for video) or directly on the original signal.
Thank you for answer
On my case, I have some data that I built them by simulation and I want to complete my database by data augmentation instead of using simulation because is laborious and time consuming
My inputs data are the parameter of building and my output is the energy consumud by the building and I use simulation software to calculate this energy.
What I want to do is to use data augmentation to vary more my parameters and calculate the energy without using ths Software .
Is this possible?
Using a generative model (such as a GAN ; generative adversarial neural network) can be helpful. It has been tried in several domains (such as [1]) and seems to be able to generate unseen and useful data points consistent with the train data distribution (training it may be harder than classical models but many implementations for different kinds of GANs can be found on github).
Other simpler techniques have also been tested in older papers (early 2000's) exploring adding noise, interpolating or extrapolating between data points to generate new ones. It all depends on the kind of data you're working on, it's distribution and the amount of variability within the dataset.
-
References :
[1] : https://www.bayeswatch.com/assets/papers/Augmenting_Image_Classifiers_using_Data_Augmentation_Generative_Adversarial_Networks.pdf
As I am not a specialist or very skilful on developping like code to training this model ( GAN ) there is any pre-developed toolbox wich can help me.
Sorry I can't help. I don't think there is any generic domain-agnostic data augmentation toolkit out there, if you're not learning directly from data (using a neural network for example), or modeling your data distribution with some generative model (GMM, HMM, etc), some domain knowledge is needed to generate new data.
I need data augmentation code for regression data (numeric values) not image.
- Badges
- Science topic