Design and Experimental Evaluation of a Voice Control System for Autonomous Robotic Agricultural Systems

Abstract

This paper presents a multilingual voice control system for autonomous agricultural machinery, addressing the need for intuitive human–machine interfaces in precision farming. The proposed architecture integrates audio capture with a configurable speech recognition module based on Faster Whisper, which transcribes commands and optionally translates from one language to another. Recognized text is processed by a hybrid natural language understanding pipeline combining intent-and-slot-filling models (BERT/T5), and a local large language model (Qwen) for commands of varying complexity. Experimental evaluation shows that medium-sized Whisper models maintain 85–90% accuracy under noise levels up to 13 dBFS, while tiny and base models degrade significantly. The Russian-to-English translation capability of medium models yields satisfactory results without fine-tuning. In the NLP stage, BERT with T5-based error correction achieves 92.7% correct command recognition at 12% Word Error Rate, while a 14B-parameter LLM reaches 96.3% accuracy at higher computational cost. Performance measurements on CPU with INT8 quantization confirm feasibility for resource-constrained edge devices. The proposed solution offers a flexible trade-off between accuracy, latency, and memory consumption, making it suitable for real-world agricultural environments with variable noise levels and mixed-language operators.