We train a neural network to detect activity of human speech in an audio frame.
The audio dataset can be found here at http://www.openslr.org/resources.php
A soft VAD value is computed and used from the speech signal and some random mixtures with chunks of noise signals for each frame of speech data.
Take Speech frames and extract Log spectrogram or MFCCs to get the features ready to be used as inputs for the neural network.
Build (a) Feedforward Model (b) sequence model ( Bidirectional LSTMs)
The model has been trained with MSE Cost function, ReLU output layer and RMSProp optimization for approximately 20-50 epochs and a dropout and decayed learning rate alongside 10% validation error.
Preprocessings involved:
- Download and extract time series waveforms from '.flac' audio files
- File-wise extract and store binary files (.mfcc)
- Mixing random frames of noise (download and extract from https://zenodo.org/record/1227121#.W23BI9VKiUm ) with clean audio signals to get time-series vectors
TODO:
- Better Hyperparameter optimization.
- Explore more complex topologies.
- Use better strategies to come up with better generalizations