A Deep Learning Approach to Speech Based Control of Unmanned Aerial Vehicles (UAVs)

摘要

Speech recognition has been one of the key research domains in computational signal processing. Despite high levels of computational complexity associated with achieving speech recognition in real-time, promising progress has been made under the umbrella of voice controlled robotics. This paper proposes an alternate approach to speech recognition for robotics applications, without adding on external hardware. We use a combination of spectrograms, MEL and MFCC features and a neural network based classification which is usually done offline, whereas the proposed method offers a remote real-time control of the robot that can be used to survey terrains that are otherwise impervious for humans, or monitor activities inside huge structures like wind-mills, gas pipelines etc. The trained model occupies lesser than 4MB on the storage medium of the platform and it also displays metrics of confidence and accuracy of prediction. The overall validation accuracy of the algorithm goes as high as 97% while the testing accuracy of the system is 95.4%. Since this is a classification algorithm, results have been presented on custom voice classification datasets.