The aim is to generate audio based on the Common Voice dataset using a Generative adversarial network.
The projects as a whole works quite good, both the generator and the discriminator are training and competing against each other. But to achieve acceptable results the generator has to be better than the discriminator, which is not he case. Even after 12 Gb of data the discriminator is still way better than the generator which basically means that the generator couldn't imitate the sound samples good enough. The expectable result is a monotonous sough. The inability of the generator to get better than the discriminator can be traced back to the data, an image(grayscale) imitating GAN for example, works with a scalar from 0-10 per pixel. One tone(compared to pixel) has 256 16bit values, with a 44 Mhz sample rate, there are a whole of 44000 * 256 * 5 values to change in a 5 second sound sample, a img generator in comparison has only 400x400 values to adopt. The complexity of the data thus would have to be reduced in either frequency or quality which both leeds to an unauthentic imitation.
> - Clone Repository
> - Install Dependencies
> - Train
> - Convert .<format> files to .wav files using tools/reformat.py <br>
> - python main.py -m train
- numpy
- Keras
- matplotlib
- librosa
- OptionParser
- uuid
- tqdm
- tensorflow
- scipy
- sklearn
- h5py
Samplerate: 44,1 kHz
Audiotype: Mono
Recommended dataset: Common Voice by Mozilla
File format: .wav
Continuous recurrent neural networks with adversarial training by
Olof Mogren Chalmers University of Technology, Sweden
-
librosa
raise NoBackendError()
audioread.NoBackendError
: install ffmpeg and
add environment variable for ffmpeg