Stream-Subs 如何在本機運作,提供無雲端 RU→EN 即時字幕
Live subtitles without the cloud: how Stream-Subs works
CJDARKLORD 開發的 Stream-Subs 是一款在本機運行的 RU→EN 即時字幕工具,透過 VAD、faster-whisper、本機翻譯和 WebSocket 將字幕送至 OBS 疊加層,無需帳戶、雲端服務或按分鐘付費。
Caption services promise a lot until you stream at 2 a.m. and the subtitles lag by three seconds, the browser tab eats 4 GB of RAM, and the monthly limit silently resets your OBS setup.
So I built a local alternative: Stream-Subs — live RU→EN subtitles that run on your own machine, with no account, no cloud and no per-minute billing.
The pipeline
mic → VAD (speech only) → faster-whisper → RU→EN translation → WebSocket → OBS overlay
A few decisions that turned out to matter:
- VAD first. Sending silence to Whisper is how you get hallucinations — subtitles for sentences nobody said. A voice-activity gate drops silence and noise before recognition.
- faster-whisper over reference Whisper. CTranslate2 backend, int8 quantization. On an RTX 2070 a 30-second fragment is processed in 0.36–1.62 s depending on the model and the amount of speech, using about 4.4 GB of VRAM for the medium int8 model.
- Local translation, not an API. That keeps latency down and the audio offline. (One trap: the NLLB model is CC BY-NC — fine for personal use, not for a commercial product. I swapped it out and wrote about that separately.)
- OBS overlay, not a separate window. The subtitle layer is a browser source, so it composites exactly like any other scene element.
- A hallucination filter after recognition. Repeated phrases, artefacts of silence and banned words are dropped before they hit the screen.
What it costs to run
Nothing per minute. It runs on your GPU (or CPU, slower). The trade-off is explicit: the quality ceiling is your hardware, and you're responsible for the models and updates. In exchange there's no subscription, no audio sent anywhere, and no service that can change its pricing next quarter.
Honest limits
- Real-time quality depends on the model size and your GPU;
- It is not a replacement for a human interpreter in a legal or medical context — it's a streaming accessibility tool;
- The first launch downloads about 4.9 GB of models; after that, starts are instant;
- Windows-only for now (the stack itself is portable).
Why open source
The code is under MIT. Subtitle tooling is exactly the kind of thing where the community's edge cases (accents, overlapping speech, game audio) matter more than a marketing page. Fork it, swap the models for your language pair, open a PR.
Project page with builds, checksums and changelog: Stream-Subs — live subtitles for streamers. The engineering story is in the case study: StreamSubs: stream subtitles without the cloud.
來源:DEV Community · dev.to