• dataset_wav --- the corpus of audio .wav files of spoken urdu words

• dataset_java_feature --- files saved with extracted MFCCs features using the librosa java implementation. MFCCs are saved in seperate text files with the same name as the audio file for each audio later to be used for comparison with the librosa python generated MFCCs.

• librosa/librosa_java --- the java implementation for librosa. Reads the audio .wav files to extract MFCCs and saves them in seperate text files for each audio.

• librosa/librosa_python --- python librosa code to extract MFCCs from .wav audio

• mfcc --- code for extracting features using python_speech_features library (wav2mfcc.py) & using librosa library (wav2mfcc_librosa.py)

• model --- code for LSTM Model for classification of native and non-native speakers (native_classifier.py)

• compare --- code to compare MFCCs generated from python librosa and java librosa using RMSE aand generate graph for visualization

• Others --- contains potential code for feature extraction using java (eg. OpenIMAJ library) from other different libraries that were tried in the study.

• utilities --- Creating Wav audio object in python ( data_loader.py )

• train_model.py --- train model using MFCCs extracted using python librosa

• train_model_librosa.py --- train model using MFCCs extracted using librosa java implementation

• sib_app --- Android application

https://www.kaggle.com/hazrat/urdu-speech-dataset

• Create a virtual environment

virtualenv env

• Install all the requirements in the virtual environment

source env/bin/activate
pip3 install -r requirements.txt

• Run the script that trains the librosa model

source env/bin/activate
python train_model_librosa.py

• Add these lines to your Python script

from keras.models import load_model

model = load_model('model.h5')
model.summary()

• Run the script that compares the two implementations

source env/bin/activate
cd compare
PYTHONPATH=../ python compare_output.py