Detects number of people using Audio

Dataset used Libra MultiSpeaker dataset

validation accuracy is over 40%. Converted audio to spectrogram with 512 fft length and used CNN network to detect spikes in spectrogram.

Problems to solve :

1. tf data microphone library outputs and spicy spectrogram scaled output validation

2. Scaling input data. Presently using tf norm layer to normalize