I used to send film clips to cloud dubbing services and watch the meter run. Five minutes of a French documentary was about a dollar. A full episode runs longer, and the numbers get stupid fast. That's the whole reason I stopped looking at the meter and started dubbing locally instead.
The short version: cloud dubbing charges by the minute, by the character, or by the month. Local dubbing costs a one-time model download and your electricity. My outage bill for a year of local video translation was roughly what that cloud service dinged me for two afternoons.
What cloud dubbing actually charges
Most cloud dubbing and translation tools bake their pricing into usage. You get a free tier that's measured in seconds, then it's per-minute for audio, per-character for generated speech. Ten hours of dubbed video a week adds up to a very specific number. Some roll it into a flat monthly sub, so if you don't use it every single day you're paying to keep it warm.
Then there's the stuff nobody prices into the comparison. Your audio goes to their servers, so the clip you're translating is now someone else's dataset. It's usually against their ToS to dub proprietary or sensitive media at all. And you need a connection for every single sentence.
Why I moved translation fully local
My whole pipeline runs on one Windows box. The browser extension grabs the tab audio, Whisper transcribes it, LM Studio runs a small local LLM to translate, and Kokoro TTS speaks the result. It's a ~2GB model, Python 3.10, no GPU required (it helps). Once set up, it works offline.
The male/female voice switching is the part that surprised people. I had a tool that always picked one voice for every speaker. Pitch analysis decides the voice per line, so a two-person scene has two actual voices, and the original audio ducks under the dub so you still hear the music and the delivery.
It's not instant. It waits a sentence or two because it can't translate until someone finishes talking. Every dubbing tool does this. The difference is the wait happens on my desk and nothing leaves the building.
The honest cost breakdown
I did the math instead of guessing, and the delta was lopsided. My monthly spend on the local rig was the model storage and fan noise. The cloud quotes I got were either a per-minute meter or a subscription that assumed heavy daily use. That subscription only made sense if I was dubbing constantly.
Privacy is the part I'd defend to a client. Proprietary training video, client interviews, unreleased content — none of it touches a third party. Disconnect the machine and it still works. You can't do that with a service that holds your audio on its servers.
FAQ
Can I dub videos in real time with a computer that has no GPU?
Yes. The bottleneck is Whisper and the TTS, and Kokoro runs fine on CPU. A slow laptop just means a slightly longer sentence gap. No GPU required.
Is everything stored locally?
Yes. After the initial model download the whole pipeline is offline. Audio, transcripts, and the translated dub never leave your machine.
What languages does it handle?
Whatever Whisper transcribes and the local LLM translates. Same coverage as the underlying models, which is a lot. I mostly dub into and out of English, Spanish, French, and Hindi.
If you want to skip the week of wiring it together, my working build is a one-time $19 at https://symshah.gumroad.com/l/livedub. Offline after setup, no meter.
Top comments (0)