NN Audio Gender Classifier

This project implement neural based classifiers for gender classification from audio files
RNN models and CNN models are implemented and compared

Setting Up Environment

Requirements needed can be found in requirements.txt
It is recommended to set up a new Anaconda environment or python virtual env and install the required requirements.

Preparing Data

Downloading Datasets

For training and validation, We have used TIMIT dataset and the 100 hr subset of LibriSpeech Dataset.
To Download The data:

TIMIT: https://goo.gl/l0sPwz
100 hr Subset LibriSpeech: https://www.openslr.org/resources/12/train-clean-100.tar.gz

Feature Extraction

Mel-frequency cepstrum (MFCCs) are calculated for each frame of the data audio files
To extract mfcc files, modify the Datasets paths and mfccs_path in config.json Output Mfccs will be saved in "mfccs_path"
Then run the command python train.py
It is expected for the paths to be passed to be structured as follow

.
├── ...
└─archive			##TIMIT_path                  
	├── PHONCODE.DOC
	├── PROMPTS.TXT
	├── README.DOC
	├── SPKRINFO.TXT
	├── SPKRSENT.TXT
	├── test_data.csv
	├── TESTSET.DOC
	├── TIMITDIC.DOC
	├── TIMITDIC.TXT
	├── train_data.csv
	└── data
		 ├── TEST
		 └── TRAIN
 ...
.
└─── LibriSpeech		#LibriSpeech Path          
	├──BOOKS.TXT
	├──CHAPTERS.TXT
	├──LICENSE.TXT
	├──README.TXT
	├──SPEAKERS.TXT
	└──train-clean-100 ...

Configurations File

Modify the needed hyper parameters in config.json file, you can modify files Sampling Rate, Learning Rate, Number of epochs, and other parameters
You can also choose the model type RNN or CNN by setting the parameter network_type
This is an example of config.json

{
    "out_path": "out/",
    "mfccs_path": "/data/datasets/mfccs",
    "TIMIT_path" : "/data/datasets/archive",
    "LIBRI_path" : "/data/datasets/train-clean-100/LibriSpeech",
    "test_model_path" : "/data/gender_classifier/out/CNN_model_3800Steps.pt",
    "SR" : 16000,
    "n_mfccs" : 20,
    "max_mfcc_len" : 250,
    "batch_size": 64,
    "data_shuffle": true,
    "train_percent": 0.8,
    "test_percent": 0.1,
    "val_percent": 0.1,
    "network_type": "CNN",
    "learning_rate": 0.001,
    "max_epochs": 100,
    "log_interval": 10,
    "save_checkpoint_path" : "out/",
    "save_checkpoint_interval" : 200
}

Training The Network

You can run training by running the script python train.py
Output to monitor losses and accuracies will be written in the terminal, and a tensorboard log out will be saved in out_path

Evaluating The Network

you can run python evaluate.py to evaluate and calculate the prediction metrics for the model on test set.
To evaluate a model, specify the "test_model_path" and "network_type" in the config.json. The test model path is the path of the model to load and run testing with

Using network to predict audio file gender

you can use a checkpoint of a trained model to genrate prediction for a file by running the command
python infer.py <audio file path>

example:

python infer.py /data/test_wavs/out.wav
You have to specify the inference model path in the config.json file you can download the best performing model from this link "https://drive.google.com/file/d/1cMynQeKceuEMHVZRf7CV2t_1cjr0cCHB/view?usp=sharing"
Then modify the config file, with the model path and "network_type" set to "CNN" before running the script

OthmanMohamed/NN_Audio_Gender_Classifier