Transcription
Whisper (Local)
Offline Whisper transcription on Windows with whisper.cpp, on the CPU or on a graphics card with NVIDIA CUDA or Vulkan.
About
Whisper (Local) adds offline Whisper transcription to TypeWhisper for Windows. It runs whisper.cpp through Whisper.net, downloads GGML model files when you ask for them and transcribes on your PC, without sending audio to a cloud service.
Features
- Offline Whisper transcription on Windows
- 14 models from Tiny to Large V3 Turbo, as multilingual, English-only and smaller Q5_0 variants
- Multilingual models cover 99 languages and detect the spoken language automatically
- Live preview while you dictate
- Your dictionary terms go to Whisper as its initial prompt, so names and technical terms are recognized more reliably
- Native translation into English with multilingual models other than Large V3 Turbo
- Processing device: Automatic, CPU, NVIDIA CUDA, Vulkan or AMD ROCm
- The Processing device setting shows what the loaded model runs on, with Vulkan including the graphics card in use, or why it fell back to the CPU
- Download, use and remove models on the add-on’s page
- Free disk space is checked before a download, and every download must match the exact size of the model file
- No API key or account required
Transcription Models
| Model | ID | Size | Languages | Translation |
|---|---|---|---|---|
| Tiny | tiny | ~75 MB | 99 | Yes |
| Tiny (English) | tiny.en | ~75 MB | English | No |
| Tiny (Q5_0) | tiny-q5_0 | ~31 MB | 99 | Yes |
| Base (recommended) | base | ~142 MB | 99 | Yes |
| Base (English) | base.en | ~142 MB | English | No |
| Base (Q5_0) (recommended) | base-q5_0 | ~57 MB | 99 | Yes |
| Small | small | ~466 MB | 99 | Yes |
| Small (English) | small.en | ~466 MB | English | No |
| Small (Q5_0) | small-q5_0 | ~182 MB | 99 | Yes |
| Medium | medium | ~1.5 GB | 99 | Yes |
| Medium (English) | medium.en | ~1.5 GB | English | No |
| Medium (Q5_0) | medium-q5_0 | ~601 MB | 99 | Yes |
| Large V3 Turbo | large-v3-turbo | ~1.6 GB | 99 | No |
| Large V3 Turbo (Q5_0) | large-v3-turbo-q5_0 | ~684 MB | 99 | No |
Processing Device
| Option | What it does |
|---|---|
| Automatic | Tries NVIDIA CUDA, then Vulkan, then the CPU. It does not download the CUDA files; choose NVIDIA CUDA once to install them. |
| CPU | Runs on every PC. |
| NVIDIA CUDA | Needs an NVIDIA graphics card. When the model loads for the first time, TypeWhisper downloads NVIDIA’s cuBLAS libraries. Then restart TypeWhisper and select the model again. |
| Vulkan | Runs on a graphics card through its Vulkan driver. The add-on picks the first dedicated graphics card and uses integrated graphics only when there is none. |
| AMD ROCm | For a ROCm build of whisper.dll that you provide. Set the environment variable TYPEWHISPER_WHISPERCPP_ROCM_LIBRARY_PATH to the file or its folder and restart TypeWhisper. |
Setup
- In TypeWhisper, open Settings > Discover plugins and install Whisper (Local). It then appears in the Settings sidebar, and its page opens.
- Under Models, select Download model next to the model you want. Base and Base (Q5_0) are marked Recommended.
- Optional: under Transcription, choose a Processing device and select Save settings. Do this before you use the model, so that it loads on this device.
- Select Use model to use it for dictation.
- Optional: to translate into English, use a multilingual model other than Large V3 Turbo and choose Translate to English as the Transcription task in the Dictation settings.
Notes
- English-only models always transcribe English. Multilingual models detect the language unless you set one under Spoken language in the Dictation settings.
- English-only models and Large V3 Turbo, including its Q5_0 variant, cannot translate. With Translate to English selected, dictation reports this instead of recording.
- The prompt with your dictionary terms is limited to 224 bytes; terms stay whole and in dictionary order. If a transcript only repeats dictionary terms, the add-on decodes the audio again without the prompt, so silence does not turn into dictionary terms.
- Once a runtime has loaded, switching to a different processing device can require restarting TypeWhisper.
- Quantized Q5_0 variants are smaller downloads than the full models. Larger models can be more accurate but need more disk space and memory.
- The add-on stores model files in its plugin data directory under
Models/. - Older versions of the add-on were called whisper.cpp (Local).