New voice AI lets you control drones with natural speech at 93% accuracy

Voice control promises an intuitive alternative to manual drone piloting, but most systems fail when faced with the spontaneous, disfluent speech of non-expert users. A new paper accepted at RO-MAN 2026 tackles this head-on by proposing an End-to-End Spoken Language Understanding architecture that processes natural French commands in real-time. The model combines a frozen Self-Supervised Learning (SSL) acoustic encoder with a lightweight LSTM-based classification head, augmented by cross-modal knowledge distillation to align acoustic representations with semantic embeddings from a text teacher—critically, no transcription is needed at inference.

On the novel VoiceStick corpus—collected from 29 nonexpert dyads during real teleoperation sessions—the best configuration achieved 93% accuracy on simple voice commands with only 7ms inference latency, a 29x speedup over cascade baselines (79% accuracy, 202ms). For the full spontaneous speech test set, accuracy reached 82%, with cross-modal distillation consistently improving robustness across configurations. These results demonstrate that end-to-end architectures are not only feasible but preferable for real-time, voice-guided drone teleoperation, combining semantic robustness with low latency and calibrated confidence.