Voicenet

Comprehensive Python library for speech 💬 and voice 🔉.

Getting Started

Voicenet is a comprehensive library for performing speech/voice based functions. It is capable of doing:

• Gender detection based on the voice.
• Feature extraction from voice

Build Status

In development, released first model.

Installation

With pip

From Source

git clone https://github.com/Robofied/Voicenet
cd Voicenet
pip install -e .

Hardware Requirements

Models are trained and run on these machines -:

CPU(Octa-Core) 16GB RAM - Ubuntu 19.04

Getting Started

Downloading pre-trained models and data

You can download the models and data manually from the GitHub releases or use our download functions:

from voicenet.datasets import stamerican

(x_train, y_train), (x_test, y_test) = stamerican()

stamerican will load data locally(relative to ~/voicenet/data/datasets)

If you are looking to work on ST-AEDS dataset with model training(we arlready have trained model but incase you want to train model again.) then you can skip this step as training module already perform this step.

Making Predictions

Trained models are saved in 'models/' directory

from voicenet.models import GenderDetection

## VoicePipeline is initialize with "gmm" model as it will use by default gmm models
## VoicePipeline(model="gmm") //by deafult, other options are not available for now. 
voicenet = GenderDetection(model="gmm")      
winner = voicenet.predict('wav-file-path.wav')

print(winner)

Preparing Your Data

If you are planning to use training modules of Voicenet then you can skip this step totally for ST-AEDS dataset and for any other dataset you have to download data in a specific format and then skip it if want to use training modules

1. ST-AEDS Dataset

a. Manual

You have to download ST-AEDS Dataset from here and then extract it and split it into training and testing sets for females and males both separately as shown below(directory structure)

b. With converters

from voicenet.datasets import stamerican

(x_train, y_train), (x_test, y_test) = stamerican()

stamerican will load data locally(relative to ~/voicenet/data/datasets)

2. Any other dataset(Custom dataset)

a. Manual

You have to download/save your data into a directory which also contains subfolders female, male with thier respective females and males voice file.

Moving forward you need to prepare data i.e, training/testing data splitting and follow this directory structure

b. With converters

You have to dowload/save your data into a directory which should contains subfolders female, male with thier respective females and males voice file.

But in order to split dataset you can use -:

from voicenet.data_preparation import SplitData

data_dir = 'path-to-load-data-from-containing-females-males-folder'

data_dir_females = 'path-to-load-data-from-containing-females-males-folder-and-load-only-females'

data_dir_males = 'path-to-load-data-from-containing-females-males-folder-and-load-only-males'

SplitData.universal_data_preparation(data_dir, data_dir_females, data_dir_males)

Training Models

1. ST-AEDS Dataset

from voicenet.training import GMMModelTraining

gmm_model = GMMModelTraining()

data_dir = 'path-to-load-data-from'
model_save_dir = 'path-to-save-trained-model'

gmm_model.train_model(data_dir = data_dir, model_dir = model_save_dir)

2. Any other dataset

from voicenet.training import GMMModelTraining

gmm_model = GMMModelTraining(staeds_flag=False)

data_dir = 'path-to-load-data-from-containing-females-males-folder'

data_dir_females = 'path-to-load-data-from-containing-females-males-folder-and-load-only-females'

data_dir_males = 'path-to-load-data-from-containing-females-males-folder-and-load-only-males'

model_save_dir = 'path-to-save-trained-model'

## Examples to load females/males data
gmm_model.train_model(data_dir='data/raw/test', data_dir_females='data/raw/test/females', data_dir_males='data/raw/test/males', model_dir='models/')

Contributing to Voicenet

All contributions, bug reports, bug fixes, documentation improvements, enhancements and ideas are welcome.

A detailed overview on how to contribute can be found in the contributing guide. There is also an overview on GitHub.

If you are simply looking to start working with the voicenet codebase, navigate to the GitHub "issues" tab and start looking through interesting issues. There are a number of issues listed under Docs and good first issue where you could start out.

You can also triage issues which may include reproducing bug reports, or asking for vital information such as version numbers or reproduction instructions.

Or maybe through using you have an idea of your own or are looking for something in the documentation and thinking ‘this can be improved’. You can do something about it!

Feel free to ask questions on the mailing list or on Slack.

License

BSD-3