ThomasH88 / subvocalization

Using spectrograms and machine learning to predict subvocalized commands

Subvocalization

Identify when to turn on and off the lights using subvocalization.

Project Description

The subject used the OpenBCI Ganglion with 4 electrodes placed on the subject's neck and chin and then proceeded to record 1000 samples of silent speech. 500 samples of "Lights-on" and 500 of "Turn-off".

The goal is to use this data to build a model that can recognize when the subject is subvocalizing "Lights-on" and "Turn-off" in order to be able to control a device e.g. a lamp, the way you would using Alexa, Siri or Google Home but without the need to articulate the commands.

Data

• The dataset contains 1000 measurements of sEMG using the OpenBCI Ganglion and 4 channels
• 500 were recorded subvocalizing "Lights-on"
• 500 were recorded subvocalizing "Turn-off"
• Special thanks to Taylor Yang who made the dataset

Process

• The data was converted to a csv file for easy importing
• The data was then converted to multiple spectrograms
• The spectrograms where then used to train a ConvNet
• Achieved 98% accuracy on the test set