LinguoAI
EN
Deutsch English Русский Українська Türkçe Español Italiano
Set up the desktop app

Step by step

Set up LinguoAI on your PC.

Here you can see what to click, what to type and how to know when you are done. Start with step 1. More options are at the bottom.

Start with step 1 Open Colab manual Find your version

Set up once

Install on your Windows PC

You need a Windows PC with internet access and several gigabytes of free space. There is no ready-made installer yet. The steps below set up the program. You do not need to know how to code.

  1. 01

    Install two helper programs

    Open the Windows Start menu, type PowerShell and open the program. Copy the following lines into it one at a time, pressing Enter after each. Wait for each installation to finish before entering the next line.

    PowerShell
    winget install --id Python.Python.3.11 --exact --source winget
    winget install --id Gyan.FFmpeg --exact --source winget

    Python runs LinguoAI. FFmpeg reads and writes videos. If Windows asks for confirmation, check that the prompt is for this installation.

    Done? Both installations report success or “already installed”. Now close PowerShell.

  2. 02

    Download and unzip LinguoAI

    Download the ZIP file. Right-click it in your Downloads folder and choose “Extract All”. Then open the extracted folder.

    Download LinguoAI 2.2.2 ↓

    You are in the right place if you can see the file pyproject.toml and the folder linguoai .

  3. 03

    Start in the right folder

    In the open folder, click the address bar at the top. Type powershell and press Enter. A new window opens. Copy these three check commands into it one at a time:

    PowerShell
    py -3.11 --version
    ffmpeg -version
    ffprobe -version

    You see Python 3.11 and two version reports. If you see “not found”, use the help below before continuing.

  4. 04

    Set up LinguoAI

    Stay in this window. Copy each line separately and press Enter. After each line, wait until you can type again. Larger files download the first time.

    PowerShell
    py -3.11 -m venv .venv
    .\.venv\Scripts\python.exe -m pip install --upgrade pip
    .\.venv\Scripts\python.exe -m pip install torch --index-url https://download.pytorch.org/whl/cpu
    .\.venv\Scripts\python.exe -m pip install -e ".[desktop]"

    Finished without an error? Setup is complete. You only need to do it once.

  5. 05

    Open the program

    Copy this line into the same window:

    PowerShell
    .\.venv\Scripts\python.exe -m linguoai

    The LinguoAI window opens. Keep PowerShell open while you use LinguoAI.

Try it now

Your first translated video

Start with a short video with clear speech, for example 30 seconds long.

  1. Under „Start“ click „Datei …“ and select your video.
  2. Select your preferred „Zielsprache“ . The source language can be detected automatically.
  3. For this first test, select “Google Translate · online” and “Edge TTS · online”.
  4. Open “Erweitert” → “Leistung”. Under “Hardware für Spracherkennung”, select “Nur CPU”. Then return to “Start”. Leave the other settings as they are.
  5. Click „Verarbeitung starten“ . The first time, LinguoAI also downloads the speech recognition model. This can take a while.
  6. When you see „FERTIG“ click „Video öffnen“ . Use „Ordner öffnen“ to find the video and both subtitle files.

If you get stuck

How do I open LinguoAI next time?

Open your extracted LinguoAI folder again. Type this in the address bar at the top: powershell. Run only the command from step 5. You do not need to install everything again.

“winget” cannot be found

Install or update Microsoft’s “App Installer” in the Microsoft Store. Then reopen PowerShell and repeat step 1.

Open Microsoft’s guide ↗
Python or FFmpeg cannot be found

Close PowerShell and open it again as in step 3. If that does not help, restart your PC and check the three lines again. This guide requires Python 3.11 to be installed.

A command reports an error

Do not run the next line yet. First check that you are in the extracted folder containing pyproject.toml and that you copied the whole line. Keep the exact error message in case you need help.

It takes a long time or the video never finishes

Models download on the first run. Keep the window open and check whether progress continues. If LinguoAI shows an error, read the message in the log. Then try a shorter video first.

More options: Linux, local voices and more
Linux

Python 3.10+ and FFmpeg required

# FFmpeg, ffprobe und Tkinter über die Paketverwaltung installieren.
# Im entpackten Projektordner:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
python -m pip install -e ".[desktop]"
ffmpeg -version
ffprobe -version
python -m linguoai

Four ways to get a voice track

Compare voices

“Own voice” can mean different things. This comparison will help you make the right choice.

Edge TTS

The easy way to get started with ready-made voices from the Microsoft catalog.

Online · No, catalog voice

gTTS

Fast Google Voice via Google Translate Text-to-Speech with voice code.

Online · No, language code

Piper

Fast local output with an existing or self-trained ONNX model.

Local · CPU, optional CUDA

Chatterbox V2

Adopts vocal character and manner of speaking from a short, clean recording.

Reference voice · GPU recommended

Add Piper locally

Install in the same virtual environment and then load a suitable voice:

.\.venv\Scripts\python.exe -m pip install -e ".[desktop,piper]"
New-Item -ItemType Directory -Force ".\Stimmen"
.\.venv\Scripts\python.exe -m piper.download_voices --download-dir ".\Stimmen" de_DE-thorsten-medium

Piper: ONNX plus matching JSON, optional speaker ID and adjustable speaking rate.

Own reference voice with Chatterbox

Due to audio/torch dependencies, a dedicated Python-3.11 environment is recommended:

py -3.11 -m venv .venv-chatterbox
.\.venv-chatterbox\Scripts\python.exe -m pip install --upgrade pip
.\.venv-chatterbox\Scripts\python.exe -m pip install -e ".[desktop,voice-clone]"
.\.venv-chatterbox\Scripts\python.exe -m linguoai
  • 6-10 seconds, one person, little reverb, no music or background noise.
  • Approximately 3.7GB model download; CPU is possible, but very slow for longer videos.
  • The native installation is experimental; Colab or Linux is more reliable for audio packet problems.

For repeatable processes

Command line

This example translates locally to NLLB and uses a Piper model. The same settings as in the GUI are available for automation.

Own output and model paths

PowerShell · Piper locally 
.\.venv\Scripts\python.exe -m linguoai `
  --input "C:\Videos\quelle.mp4" `
  --output "C:\Videos\quelle.de.mp4" `
  --language de `
  --service nllb `
  --device cpu `
  --nllb-device cpu `
  --tts-engine piper `
  --tts-device cpu `
  --piper-model "C:\Stimmen\de_DE-thorsten-medium.onnx" `
  --piper-config "C:\Stimmen\de_DE-thorsten-medium.onnx.json"

View all options: .\.venv\Scripts\python.exe -m linguoai --help

Python-Wheel

The finished package for an existing Python environment; FFmpeg is still required.

Load Wheel ↓

python -m pip install ".\linguoai-2.2.2-py3-none-any.whl[desktop]"
python -m linguoai

Distribute computing power sensibly

CPU and GPU are not an all-or-nothing decision.

Whisper, NLLB and the local voice can be placed separately. So you can start small and only put the part on the GPU that really benefits from it.

“Automatic” checks CUDA and otherwise falls back to CPU.

Important: Selecting “GPU” does not automatically install CUDA.
  • NLLB & Chatterbox

    Need a matching CUDA edition of PyTorch.

    PyTorch selection ↗
  • Faster-Whisper

    Additionally requires CUDA-12-cuBLAS and cuDNN 9 in the system path.

  • Piper

    Uses ONNX Runtime: onnxruntime-gpu install instead of the CPU edition.

Optional faster recognition on the CPU

Parakeet or Whisper?

Parakeet · 25 languages

NVIDIA Parakeet v3 recognizes 25 European languages. The optional sherpa-onnx adapter runs locally on the CPU with four threads by default.

NVIDIA model card ↗

Whisper · versatile

Faster-Whisper remains the default, with automatic language detection, a broader language selection, and CPU or GPU processing.

Faster-Whisper ↗

The primeline fine-tune specializes in German. Use the German profile with a compatible ONNX export for it; choose the NVIDIA profile for multiple European languages. A speed or quality advantage is not guaranteed without comparison using your own files.

Set up Parakeet

Install the optional package and provide a compatible model folder. Models are not downloaded automatically.

.\.venv\Scripts\python.exe -m pip install -e ".[parakeet]"
.\.venv\Scripts\python.exe -m linguoai --help

Use the complete sherpa-onnx model folder with encoder, decoder, joiner, and token file. For the German export, also keep the external encoder file if it is part of the model.

In the desktop app and Colab, select the speech recognition engine, profile, and model folder. Specify the source language explicitly so that translation receives the correct language code. Parakeet transcribes; the selected translation and voice follow afterwards.

sherpa-onnx models and setup ↗ · primeline model card ↗

Upload → processing → download

Work in the browser.

Video Studio

For users who have received Studio credentials. Sign in and upload videos directly in your browser.

Processing runs on the server. Translation is still limited; the full workflow has not yet been confirmed.

No Google account needed. If you do not have Studio credentials, you can choose Google Colab or the desktop version.

Open Video Studio Find your version

Google Colab

The Colab web interface already accepts uploads and Google Drive file paths. It processes videos on the active Colab machine and offers the video and subtitles for download. You can also save the results to your Drive.

Open Colab manual Open the Notebook directly

The complete toolbox

All functions in detail.

Files and orders
  • Process a single video or all videos directly contained in a folder.
  • Inputs: MP4, MKV, MOV, AVI, WEBM and M4V; Editions: MP4, MKV or MOV.
  • Automatic output suggestion, collision-free names and controlled overwriting.
  • Progress per file, current work step, cancellation and technical log.
  • Single and batch processing: upload multiple videos or enter MyDrive file paths, one per line. Upload Piper files and a reference recording, or provide their MyDrive paths.
Transcription with Faster-Whisper
  • Models of tiny until large-v3, additionally own model values.
  • Automatically detect source language or set as language code.
  • CPU, NVIDIA-CUDA or automatic selection; int8, float16 and float32 configurable.
  • Overlapping audio sections, true time offsets and duplicate segment cleanup.
  • Chunk lengths, overlap and worker number via advanced options or CLI.
Translation and text protection
  • NLLB-200 local, Google Translate without key, DeepL-SDK or Gemini.
  • Optional fallback only after explicit selection - otherwise only one provider will be contacted.
  • URLs, email addresses, currencies, units and percentages are protected during translation.
  • Timeouts, limited repeats and cache for identical text sections.
  • Optionally, Gemini corrects the Whisper transcript without changing segment IDs or times.
  • Own NLLB compatible models, fixed revisions, FLORES codes, batch and token limits.

The default NLLB-600M model is under CC-BY-NC-4.0, is a research model and not a certified production translation. Always check results.

Timing, tone and export
  • Normalize each clip to 48 kHz mono PCM and position it with sample accuracy.
  • Speed up speech that is too long within a configurable limit.
  • Warning when accelerating quickly; controlled error instead of incomprehensible output.
  • Exact end duration, adopted metadata/chapters and, if possible, unchanged video stream.
  • Atomic result group from new video, original SRT and translation SRT.
Where your data is processed
  • NLLB · Piper · Chatterbox

    After downloading in your chosen term. In Colab that means inside the temporary Google VM.

  • Google · DeepL · Gemini

    Only receive text sections when selected or as an explicitly activated fallback.

  • Edge TTS · gTTS

    Send each translated cue to be dubbed to the respective online language service.

  • Protected Gradio web interface with random password per session.
  • Optional model cache and atomic result copy in Google Drive.
  • Drive paths remain within MyDrive; Drive is not shared as a public download folder.

Temporary results will not be published until video and both subtitles are complete. Abort, limited retries, protected destination paths and redacted API keys avoid unpleasant surprises.

OPTIONAL

Your own address with Cloudflare Tunnel.

With SHARE_PROVIDER = "gradio", you automatically get a dynamic link. Google Drive is optional.

Optional Cloudflare

Only for your own fixed address: configure a named Cloudflare Tunnel with your domain and the service http://127.0.0.1:7860. This is not a publicly provided LinguoAI server.

In the Notebook, set SHARE_PROVIDER to "cloudflare", enter your hostname in CLOUDFLARE_TUNNEL_HOSTNAME and store CLOUDFLARE_TUNNEL_TOKEN as a Colab Secret. Never publish the token.

Cloudflare: 95 MB per upload by default. Gradio: up to 5 GB per file. Use Google Drive for large videos. Temporary trycloudflare.com links are not used because they do not support SSE.

Open Colab manual