How to Dub Your Own Voice Into 23 Languages With Free Local AI, and Catch the Bad Lines Before Anyone Hears Them
Cloning your voice is the easy part now. Knowing whether the Japanese version actually says what you wrote is the hard part. A free local tool called EchoLocal builds the check into the workflow. Here is the routine, and the license fine print you need to read first.
AI voice cloning has a quality problem nobody talks about, and it is not the voice.
The voices are good. Give a modern model ten seconds of you talking and it will read a script back in your voice, in Spanish, in German, in Korean, convincingly enough that a stranger would not blink. The problem sits one step later. You do not speak Korean. So when the model swallows a word, repeats a syllable, or reads a number as something else entirely, you have no way to hear it. You publish, and a viewer in Seoul hears your voice say something slightly broken.
That is the real cost of multilingual dubbing for a solo creator. Generation got cheap. Checking stayed impossible.
A free tool called EchoLocal, posted to GitHub on November 19, takes a simple and smart swing at the checking problem. Every clip it generates is transcribed back to text by a second model and compared against what you wrote. The words that do not match land in a short report. Instead of listening to 23 language files hoping to catch something, you listen to the handful of words the report points at.
My position: the read-back check is the part worth stealing, whether or not you end up using this exact tool. And this exact tool comes with a license that rules out paid work, so read that section before you build anything around it.
What the workflow is
EchoLocal ties three free models together and runs them on your own computer. Nothing is uploaded once setup finishes.
- Chatterbox Multilingual, from Resemble AI, does the voice cloning and speaking. It handles 23 languages, from Arabic and Chinese to Swahili and Turkish.
- NLLB-200, from Meta, makes a first-draft translation of your script, if you want one.
- Whisper, the speech-to-text model, does two jobs: it turns recordings into subtitles, and it listens to every generated clip and writes down what it heard.
The routine is: write your script once, translate it into drafts, fix the drafts, generate, then let Whisper read everything back and tell you where to listen.
Why the check works
It helps to know what is going on under the hood, because it tells you how much to trust the report.
When the voice model reads a line, it can go wrong in a few typical ways. It can drop a word, garble a rare one, run two words together, or drift into mumbling on long sentences. Whisper then transcribes that audio as text, without seeing your script. If the result matches the script, the clip probably says what you wrote. If it does not match, something is worth a listen.
That last word, "probably," matters. EchoLocal's own documentation puts it well: "A Whisper mismatch is something to listen to, not proof of an error." Whisper makes its own mistakes on rare words, poetic phrasing, some inflections and accent marks. So the report gives you two kinds of flags mixed together: real problems in the voice, and mishearings by the checker. You still need ears, ideally a native speaker's, for anything that matters. The difference is you are now listening to six flagged words instead of four minutes of audio per language.
The tool also does something that heads off many errors in the first place. It generates each sentence on its own, then joins them if you ask for one file per language. Long text makes Chatterbox degrade, so the script is fed in one line at a time. That one design choice is why the workflow asks you to write one sentence per line.
Put this into practice
Here is a full session. Plan on an evening the first time, mostly for setup and downloads.
1. Check your machine. On Windows, an NVIDIA graphics card with 8 GB of memory is enough, according to the README. Apple Silicon Macs work. Without a supported graphics card it runs on the processor, at about a minute per sentence, which is fine for a short intro and painful for anything long. You need about 11 GB of disk space for the models and software.
2. Install. Download the repo. On Windows, run setup.ps1 from PowerShell; on a Mac, run bash setup.sh. Setup builds its own isolated software environments and downloads the three models (about 8 GB in total). After that, it works with no internet connection.
If you plan to write every translation yourself, you can skip the translation model during setup (-SkipTranslation on Windows, --skip-translation on Mac). Keep that in mind for the license section below.
3. Record your voice sample. This step decides more than any setting. Record 5 to 15 seconds of you speaking naturally, one voice, no music. Only the first 10 seconds are used. Here is the line from the docs every podcaster should tape to their monitor: "Microphone and room end up in the cloned voice, so they matter as much as the speaker." Record in the same space and with the same mic as your real show, or the dub will sound like a different room. Drop the file in the voices/ folder or record it in the web interface; the file name becomes the voice name.
4. Write the script one sentence per line. Short, plain sentences translate better and generate better. Spell out numbers you care about. Avoid idioms if you are going to use machine translation, because the docs say NLLB-200 "is good on plain sentences and weak on poetry or idioms."
5. Translate, then actually read the drafts. In the web interface, pick your voice, paste the script, tick the languages and press Translate. You get an editable box per language. This is where a bilingual friend, a client contact or a careful pass with a second translation tool pays off. Fix the boxes before you generate. Generating a mistranslation in a perfect voice is the worst outcome, because it sounds trustworthy.
6. Generate. Press Generate. Choose one file per language, with lines joined by a short pause, or one clip per line if you are going to edit them in your timeline. Each language appears as soon as it is done, so you can listen to the first while the rest render. If your delivery sounds flat, the command-line version has an expressiveness setting (--exaggeration, from 0.25 to 1.0, default 0.5) worth testing on a single line before you run the batch.
7. Read the report and listen where it points. Open check_report.txt. For each flagged word, play that line. Decide: is the voice wrong, or did Whisper mishear? Regenerate only the lines that are actually broken. Then, for anything public, get one native listener per language to hear the final file. The report shortens that ask from "please review this" to "please check these three lines."
8. Bonus: subtitles from the same tool. The speech-to-text tab turns any recording into a transcript plus an SRT subtitle file with timing. If you are dubbing a video, that gives you subtitles in the source language for free.
The license problem, said plainly
This is the part that decides whether the workflow fits your channel.
EchoLocal's own code is released under the PolyForm Noncommercial License. In plain terms: free for personal, research, educational and other non-commercial use. Its translation model, NLLB-200, is also non-commercial only, under CC BY-NC 4.0.
The other two pieces are more open. Chatterbox and the Whisper version it uses are both MIT licensed, which allows commercial use. So if you dub client work, sponsored videos or anything you are paid for, EchoLocal as shipped is the wrong tool, and skipping the translation model does not fix that, because the app's own code carries the non-commercial license. The commercial route is to use Chatterbox and Whisper directly, or in a tool whose license allows it, and copy the routine: one sentence per line, read-back check, listen where flagged. The method is not owned by anyone.
Whether a monetized channel counts as commercial is a legal question, and I am not a lawyer. If money changes hands, assume it does and read the license text yourself.
Where it breaks
The checker is not a translator. Whisper confirms that the audio says the words in the box. It cannot tell you the words in the box are a good translation. Bad translation plus clean audio passes the check.
Rare words, names and brand terms. Both models stumble here, so the report will flag them whether or not they sound right. Expect to listen to your product name in every language.
Your room travels. A noisy or echoey sample becomes a noisy or echoey voice in 23 languages.
Performance, not just pronunciation. A line can be word-perfect and still land with the wrong emphasis or emotion. The check cannot hear that. You can.
Consent. The docs are explicit: clone only your own voice or voices you have clear permission to use. Chatterbox embeds an inaudible watermark in every clip so generated audio can be identified later. Treat that as a feature.
It is new. Two days old, one developer. It supports Windows and Apple Silicon Macs (where the read-back check runs on the processor, so it is slower); the web interface on Intel Macs is listed as untested.
The part to keep
The voices will keep getting better, and every new model will make the same pitch: your voice, any language, instantly. The bottleneck has moved to something no model release fixes on its own, which is knowing whether what came out is right in a language you cannot judge.
A read-back check is a cheap, honest answer to that. Try it on your next intro, in two languages you do not speak. Look at what gets flagged. Then decide which of those lines you would have shipped without knowing.