Hyeon Kyu Lee : [email protected]

Coworker : https://github.com/Jihackstory (email : [email protected])

This project is Pytorch version of this project : Voice Activity Detection based on Deep Learning & TensorFlow

To detect real-time voice based on korean speech, this project was utilized in 2020 Artificial Intelligence (AI) Grand Challenge Speech Recognition(Track2) (http://www.ai-challenge.kr/sub0101).

The main purpose of this competition is to design deep learning algorithms that detect five class of korean violence speech(threat, extortion, workplace harassment, out-of-work harassment, and nonviolence). In our secenario, we used this project to improve the performance of Speech to Text (STT).

The project consists of three modes: creating_data, training and prediction.

We used the LibriSpeech ASR corpus dataset(https://openslr.org/12/) at sampling rate 16kHz consist of approximately 1000 hours and English speech from audiobook.

The test-clean in website was used as training dataset in our project and the label can download here provided by the reference project.

In other words, in this project, the english voice dataset was used in the training step, and the korean voice dataset was used in the test step.

As mentioned above, this project is comprised of three modes : creating_data, training and prediction.

Run script (Command Line)

python main.py --mode prediction
# if you want to run training mode : --mode training

This mode is used to create input image data including the four features(MFCC, MFCC-△, MFCC-△2, RMSE of mel-spectogram) of VAD model.\

The parameter lists and defaults of this mode

- raw_data : raw data (e.g. LibriSpeech ASR corpus dataset)
- labels : labels of voice and non-voice
- save_dir : the file for saving the input images
- test_img_ratio : the ratio of test data from total input data
- parallel : using the cpu parallel processing, 'True' and 'False'

- path : the path to input dataset
- model_path : pretrained model directory
- batch_size : batch size of training (?, 16, 65)
- epochs : Nb of epochs for training
- learning_rate : learning rate
- n_cnn_filters : 32-64-128
- n_cnn_kernels : 8-5-3
- n_fc_units : 2048-2048
- n_classes : Nb of classes
- train_val_ratio : ratio dividing Nb of training dataset and Nb of validation dataset
- baseline_val_loss : early stopping parameter to stop the training

- test_data : path where to find test speech data
- model_path : pretrained mode directory
- smoothing : apply smoothing feature
- visualize : visualization 
- parallel : using the cpu parallel processing
- fig_path : save the visualized figure

• MIT license