PYTHON WORLD VOCODER:

This is a line-by-line implementation of WORLD vocoder (Matlab, C++) in python. It supports python 3.0 and later.

For technical detail, please check the website.

INSTALATION

Python WORLD uses the following dependencies:

numpy, scipy
matplotlib
numba
simpleaudio (just for demonstration)

Install python dependencies:

pip install -r requirements.txt

Or import the project with PyCharm and open requirements.txt in PyCharm.
It will ask to install the missing libraries by itself.

EXAMPLE

The easiest way to run those examples is to import the Python-WORLD folder into PyCharm.

In example/prodosy.py, there is an example of analysis/modification/synthesis with WORLD vocoder.
It has some examples of pitch, duration, spectrum modification.

First, we read an audio file:

python
from scipy.io.wavfile import read as wavread
fs, x_int16 = wavread(wav_path)
x = x_int16 / (2 ** 15 - 1) # to float

Then, we declare a vocoder and encode the audio file:

python
from world import main
vocoder = main.World()
# analysis
dat = vocoder.encode(fs, x, f0_method='harvest')

in which, fs is sampling frequency and x is the speech signal.

The dat is a dictionary object that contains pitch, magnitude spectrum, and aperiodicity.

We can scale the pitch:

python
dat = vocoder.scale_pitch(dat, 1.5)

Be careful when you scale the pich because there is upper limit and lower limit.

We can make speech faster or slower:

python
dat = vocoder.scale_duration(dat, 2)

In test/speed.py, we estimate the time of analysis.

To use d4c_requiem analysis and requiem_synthesis in WORLD version 0.2.2, set the variable is_requiem=True:

python
# requiem analysis
dat = vocoder.encode(fs, x, f0_method='harvest', is_requiem=True)

To extract log-filterbanks, MCEP-40, VAE-12 as described in the paper Using a Manifold Vocoder for Spectral Voice and Style Conversion, check test/spectralFeatures.py. You need Keras 2.2.4 and TensorFlow 1.14.0 to extract VAE-12.
Check out speech samples

NOTE:

The vocoder use pitch-synchronous analysis, the size of each window is determined by fundamental frequency F0. The centers of the windows are equally spaced with the distance of frame_period ms.
The Fourier transform size (fft_size) is determined automatically using sampling frequency and the lowest value of F0 f0_floor.
When you want to specify your own fft_size, you have to use f0_floor = 3.0 * fs / fft_size.
If you decrease fft_size, the f0_floor increases. But, a high f0_floor might be not good for the analysis of male voices.
The F0 analysis Harvest is the slowest one. It's speeded up using numba and python multiprocessing. The more cores you have, the faster it can become. However, you can use your own F0 analysis. In our case, we support 3 F0 analysis: DIO, HARVEST, and SWIPE'

CITATION:

If you find the code helpful and want to cite it, please use:

Dinh, T., Kain, A., & Tjaden, K. (2019). Using a manifold vocoder for spectral voice and style conversion. Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, 2019-September, 1388-1392.

CONTACT US

Post your questions, suggestions, and discussions to GitHub Issues.