Abstract
This paper proposes an architecture for transmitting voice over low data-rate and long range networks, such as LoRa networks. Unlike other works in the literature which employ audio compression and transmission of binary audio data, the proposed system comprises speech-to-text (STT), optional semantic summarization by using a large language model (LLM), and text-to-speech (TTS) modules to reduce the transmission time, and thus, the end-to-end latency. The required hardware and software components of such architecture are described along with their implications on performance metrics. Through the development of a proof-of-concept based on the Raspberry Pi 4 platform, we demonstrate that the proposed pipeline provides an effective communication with reduced latency. More specifically, the evaluation results reveal a significant latency reduction compared to the compressed audio approach of up to 94%. Even if no semantic compression is employed for short texts, a 50% reduction of the transmission latency can be achieved.