Voxxwire: Designing a Privacy-Preserving, Fully Offline Speech Translation System for Real-Time Cross-Lingual Communication

抄録

Multilingual communication in virtual meetings has become a cornerstone of modern global collaboration, yet the tools most people rely on for speech translation still depend heavily on cloud infrastructure. This creates a troubling combination of privacy exposure, network dependency, and recurring costs that limits their adoption in sensitive or resource-constrained environments. In this paper, we introduce Voxxwire, an open-source desktop application that performs complete end-to-end speech translation entirely on the user’s local machine—without sending a single byte of data over the internet.
Our system brings together four neural components within a unified asynchronous pipeline: Silero-based voice activity detection for segmenting speech, CTranslate2-accelerated Whisper inference for multilingual recognition, Argos Translate for offline neural machine translation across more than 49 language directions, and Piper for VITS-based speech synthesis. A key innovation is the dual-channel architecture that simultaneously processes both microphone input and system audio loopback, enabling bidirectional translation during live video-conference sessions. To tackle the re-translation feedback loop that inevitably arises when synthesized audio gets recaptured by the loopback channel, we devised a deterministic timing gate that suppresses capture during playback windows—a lightweight alternative to computationally expensive acoustic echo cancellation.