Evaluation of approaches for the recognition of semantically similar voice-commands based on public data

Abstract

The new focus of the international community to return to Moon [1] and Mars has sparked the need for new ideas of future manned missions. Unmanned vehicles are being considered to aid astronaut activities on extraterrestrial surfaces [2]. This paper is exploring the concept of controlling unmanned ground vehicles (UGV) with voice commands by evaluating existing technologies; specifically, DeepSpeech (DS) [3] and its pre-trained v0.6.0 model. The evaluation entails the capability to transcribe spoken English and distinguish edge cases of discrete commands. It will be discussed in depth how the model behaves for various categories of edge cases, including anagrams, rhymes, default commands, and coincidental inclusions. Moreover, technologies like natural language processing and machine learning will be discussed regarding functionality and benefits for command recognition in post-processing as enhancements to plain speech to text conversion for building a voice command interface for an unmanned ground vehicle.