Year: 2025
Type: article
Abstract: This study targets low-latency on-device voice control on Raspberry Pi for a Mecanum 4WD robot.Key challenges include latency jitter, ASR accuracy under mild-moderate noise, and NLU compute load.We propose a fully local pipeline: Audio → WebRTC VAD → grammarconstrained (WFST) Vosk ASR → TinyBERT NLU distilled, INT8-quantized, and lightly pruned; complemented by a confidence gate, temperature scaling, a 1-360 numeral parser, and runtime tuning (threading, non-blocking queues, CPU affinity).Experiments on RPi 5 reduce end-to-end latency, improve ASR RTF and WER/CER, and lower CPU/memory and NLU latency without material loss in intent/slot accuracy, demonstrating resource-efficient real-time voice control with privacy preserved at the edge.
Institution: Universitas Kristen Immanuel
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7128971286