Transcribe voice messages and other audio locally with faster-whisper and translate them into 20+ languages with your own Ollama model. No cloud, no API keys, no admin rights. After the first model download everything works offline.
- Speech-to-text with Whisper (
large-v3-turboby default), on an NVIDIA GPU or on the CPU - Automatic fallback to CPU when the GPU is not usable
- Translation into 23 languages through any Ollama model, or transcription only
- Auto-detects the spoken language, or you pick it yourself
- Reads almost any audio format directly:
.ogg/.opus(WhatsApp, Telegram),.mp3,.m4a,.wav,.flac, ... - Copy or save the original and the translation as a text file
- Windows 10/11 (the scripts are for Windows; the Python code itself also runs on Linux and macOS)
- Python 3.10 – 3.14 from python.org. Tick Add python.exe to PATH during installation.
- Ollama with a multilingual model, for example:
ollama pull qwen2.5:7bqwen2.5:14bgives better translations if you have the VRAM/RAM for it. Ollama may also run on another machine in your network. - Optional: an NVIDIA GPU with a recent driver. No separate CUDA Toolkit is needed. Without a GPU the app uses the CPU, which is slower but works fine for short messages.
- Download this repository (Code → Download ZIP) and extract it to a folder you can write to, e.g.
C:\Users\<you>\LocalVoiceTranslator. - Double-click
setup.bat. It creates a virtual environmentvenvand installs the dependencies, including the CUDA libraries for the GPU (~1.5 GB download). - Optional: right-click
create-shortcut.ps1→ Run with PowerShell to get a desktop shortcut.
- Make sure Ollama is running.
- Start the app with
start.bat(or the desktop shortcut). - Choose an audio file, pick the source language (or Auto-detect) and the target language, and click Transcribe & Translate.
The first run downloads the Whisper model (~1.6 GB for large-v3-turbo, ~3 GB for large-v3) to %USERPROFILE%\.cache\huggingface, or to the Model folder you chose.
All settings are saved in localvoicetranslator_config.json next to app.py. See localvoicetranslator_config.example.json for all keys and defaults.
| Setting | Description |
|---|---|
| Whisper model | large-v3-turbo = fast and accurate (default). large-v3 = slightly better, slower. small/medium for weak hardware. |
| Device | auto picks the GPU when available. cpu forces the CPU. |
| Precision | auto is fine in most cases. On the CPU int8 is fastest. |
| Model folder | Empty = default Hugging Face cache. |
| Skip silence (VAD) | Skips silent parts: faster and fewer hallucinations. |
| Source language | Auto-detect, or force a language (more reliable for very short clips). |
| Translate to | Target language, or None (transcribe only). If the audio is already in the target language, Ollama is skipped. |
| Ollama URL / model | Default http://localhost:11434 and qwen2.5:7b. |
The status bar shows "GPU error ... falling back to CPU".
The GPU could not be used; the app continues on the CPU for the rest of the session. Update your NVIDIA driver and run setup.bat again. On RTX 50-series (Blackwell) cards, if int8 fails with CUBLAS_STATUS_NOT_SUPPORTED, set Precision to float16.
"Translation failed: Ollama not reachable".
Start Ollama, check the URL, and check that the model name exists (ollama list).
"Importing the numpy C-extensions failed" after upgrading Python.
The venv still contains packages built for your old Python version. Close the app, delete the venv folder and run setup.bat again. If Windows refuses to delete files, the app is still running in the background: end pythonw.exe in Task Manager first.
The translation is poor.
Try a larger Ollama model (qwen2.5:14b, gemma3:12b, ...) and force the source language instead of auto-detect.
Whisper has a built-in translate task, but it only translates into English, and large-v3-turbo was trained without translation data and simply ignores it. One fast Whisper pass followed by an LLM translation is quicker, works for any target language, and usually reads more naturally.
localvoicetranslator/
config.py constants, defaults, load/save settings
transcription.py faster-whisper, model cache, CPU fallback, CUDA DLL paths
translation.py Ollama /api/generate
pipeline.py transcribe -> translate flow (no tkinter, testable)
gui.py tkinter interface
app.py entry point
tests/ pytest with mocked WhisperModel and Ollama HTTP calls
The tests need no GPU, Ollama or faster-whisper:
python -m pip install -r requirements-dev.txt
python -m pytest
The GUI tests are skipped when tkinter or a display is not available.
nvidia-cudnn-cu12 in requirements.txt is pinned to exactly the cuDNN version ctranslate2 was built against. When upgrading ctranslate2, check that pin as well.