A self-hosted, open-source (AGPL) real-time speech-to-speech translation platform for organizations that need multilingual communication without sending audio to the cloud. It captures live audio from microphones, meetings, browser tabs, or streams, then runs a local pipeline: Whisper for speech recognition, NLLB/Ollama-compatible models for translation, and Piper for text-to-speech. It supports 40+ languages (including Chinese, English, Japanese, Hindi, Arabic) and claims sub-second to ~2-second latency depending on hardware. Because everything runs on your own infrastructure, no audio, transcripts, or translated speech leave your servers—designed for enterprise, government, legal, and NGO use cases where data sovereignty matters. Critical caveat: This is not a SaaS product. You must deploy and maintain it yourself (Docker, models, tuning), and translation quality depends on the models you choose and the hardware you run it on.
Pros & Cons:
Pros: The self-hosted, AGPL-licensed model is the core differentiator—organizations with strict data-residency or privacy requirements can run real-time translation entirely on their own infrastructure, with zero external API calls; it covers 40+ languages including major Asian and European languages, with experimental support for more; the pipeline uses proven open-source components (Faster Whisper, Ollama/NLLB, Piper), so you can swap models or tune parameters for latency vs. quality; the GitHub repo includes benchmark scripts and detailed tuning guidance for STT, translation, and TTS stages.
Cons: It is not a ready-to-use product—you need Docker, server administration skills, and hardware (CPU or GPU) to deploy and maintain it; translation quality is model-dependent, so you cannot expect DeepL-level output unless you invest in model selection and tuning; latency claims depend heavily on hardware—the “42ms” figure is an internal metric, and real-world conversational latency may be higher on modest hardware; no official SaaS or managed hosting surfaced, so non-technical teams must either hire help or look elsewhere.
Who is it for?:Enterprises, government agencies, legal teams, and NGOs that need real-time multilingual communication but cannot send audio to external cloud services; developers and sysadmins comfortable with Docker and self-hosting who want to deploy a speech-to-speech pipeline on their own hardware; privacy-sensitive teams exploring open-source alternatives to commercial translation APIs. Not suited for users who want a “just works” web app or mobile app (no managed service), non-technical teams without server administration capacity, or anyone expecting DeepL-level translation quality out of the box without model tuning.