Voice-Controlled Vehicle: Determining the Best Machine Learning Models for Voice Recognition

Technology that helps visually impaired people often use mobility canes to navigate. Assistive navigation technologies also often rely on voice-based control systems, which can be vulnerable to unintended activations in shared environments. This research develops a one- and two-stage machine learning voice model to determine what the most accurate one is using tests on a miniature vehicle. The models were trained on a google dataset as well as user voice recordings to determine model coefficients. The models take raw wav files as input and employ a convolutional neural network to classify spoken directional commands like “go”, “down”, “left”, and “right”. In the two-stage model, the audio input gets classified as user’s audio or not user’s audio and the command is ignored if it isn’t the intended user’s audio. If the command is identified as the intended user’s voice, then it goes to the second stage for command classification. This was just an addition to see how one stage and two stages compare. It was found that the two-stage recognition maintained greater accuracy in command recognition and better resistance to command hijacking from unintended users. These results suggest that multi-stage classification outperforms single-stage voice classification when used strategically.