Cloud translation costs me a rent and a delay. My local dubber skipped both.
I watch a fair amount of Korean and Japanese stuff, and for a while I did the obvious thing: rented a cloud service to translate it as it played. It worked, sort of. But it also cost me a monthly bill, made me keep an eye on network latency, and quietly shipped every scrap of dialogue audio up to someone else's server.
So I rebuilt the same pipeline to run entirely on the machine in front of me. That gap became the whole point of the tool.
The setup is simpler than I'd have guessed. A browser extension pulls the tab audio. Whisper transcribes it in pieces, a small model inside LM Studio translates the text, and a neural TTS (Kokoro) speaks it while the original audio ducks underneath so you still hear the scene. Nothing has to leave your PC. No GPU required (it helps). Python 3.10+, LM Studio, and a translation model around 2GB.
People always ask about the delay, so I'll be straight about it. There's a natural lag of a sentence or two. That's not me being lazy. No dubbing tool on the planet can start translating a line until the speaker finishes it, because the translation depends on what the whole sentence means. The cloud has that same delay, plus a round trip on top of it. The local version waits the same sentence boundary, then starts speaking almost immediately. It ends up closer to real-time, and it doesn't cost a thing per minute.
The part that surprised me was voices. Whisper gives me the spoken text, and if I measure the pitch of the source speaker, the TTS can pick a lower register for a male voice and a higher one for a female, automatically. It's one small threshold in code, and it drags the whole thing out of robotic territory. A cloud service would've billed me per line for that.
The economics are what really pushed me local. Cloud translation isn't something you buy once; it's a rent you keep paying, and it only ever gets faster when you pay a pricier tier. A local translator is a one-time purchase that runs on hardware I already own, works offline after setup, and never counts characters. For someone who watches foreign films a few times a week, the math stopped being close.
If you watch a lot of subtitled cinema and you're tired of sending every second of audio to a server, I built a small Windows app plus a browser extension that does the whole chain locally, tab audio to Whisper to a local LLM to a natural voice. It's nineteen dollars. https://symshah.gumroad.com/l/livedub
Top comments (0)