This project contains some solutions for programming specotgram functions using keras layers. The main purpose was to overcome early feature extraction in speech and speech spoofing assignments. We mimic scipy.signal using hamming and some other implemtations In more details:
pre_emp.py:
It suggests two methdologies for perfroming pre emp to mono signal : 1 Using a regular dense layer (very "keras") 2 By regular subtraction
Kapre
We aim to mimic kapre package using keras generic layers https://github.com/keunwoochoi/kapre/. The motivation for this step is our need to perfrom time_frequency spectogram in Keras.Net We thus needed to use generic layers (keras.net is not familiar with Kapre) The output of the file is a spectogram that one can connect to a regular Conv2 Layer
scip_spectogram_keras
We implement in keras specotgram using hamming window as scipy.signal perfroms stft with Hamming