73 lines
2.8 KiB
Markdown
73 lines
2.8 KiB
Markdown
# Local Models
|
|
|
|
Blitztext can run transcription locally with WhisperKit/CoreML. The app does not bundle a speech model, but it can download the selected compatible model from Hugging Face into the local cache.
|
|
|
|
## Recommended First Model
|
|
|
|
Use Whisper Small for the first local test. It is multilingual, supports German, and is much lighter than the large variants.
|
|
|
|
- [argmaxinc/whisperkit-coreml: openai_whisper-small_216MB](https://huggingface.co/argmaxinc/whisperkit-coreml/tree/main/openai_whisper-small_216MB)
|
|
|
|
Local cache path:
|
|
|
|
```text
|
|
~/Library/Application Support/Blitztext/models/whisperkit/openai_whisper-small_216MB
|
|
```
|
|
|
|
## Other Compatible Models
|
|
|
|
You can also install larger WhisperKit CoreML models into the same cache directory:
|
|
|
|
- [openai_whisper-large-v3-v20240930_626MB](https://huggingface.co/argmaxinc/whisperkit-coreml/tree/main/openai_whisper-large-v3-v20240930_626MB)
|
|
- [openai_whisper-large-v3-v20240930_turbo_632MB](https://huggingface.co/argmaxinc/whisperkit-coreml/tree/main/openai_whisper-large-v3-v20240930_turbo_632MB)
|
|
|
|
The app detects installed model folders that contain `AudioEncoder.mlmodelc`, `MelSpectrogram.mlmodelc`, and `TextDecoder.mlmodelc`.
|
|
|
|
## Install From The App
|
|
|
|
Open Blitztext, go to **Settings > Anpassen**, choose a local model, and click **Installieren**. You can also switch on **Sicherer Lokaler Modus** from the main popover; if the selected model is missing, Blitztext starts the download and installs it into the local cache.
|
|
|
|
After the model is installed, the Blitztext transcription workflow can run in local mode. The rewriting workflows still use OpenAI, so they are paused while secure local mode is active.
|
|
|
|
## Optional Manual Install
|
|
|
|
If you prefer the CLI path, install the Hugging Face CLI so the `hf` command is available:
|
|
|
|
```bash
|
|
python3 -m pip install --upgrade "huggingface_hub[cli]"
|
|
```
|
|
|
|
Create the local model cache:
|
|
|
|
```bash
|
|
mkdir -p "$HOME/Library/Application Support/Blitztext/models/whisperkit"
|
|
```
|
|
|
|
Download the recommended first model:
|
|
|
|
```bash
|
|
hf download argmaxinc/whisperkit-coreml \
|
|
--include 'openai_whisper-small_216MB/*' \
|
|
--local-dir "$HOME/Library/Application Support/Blitztext/models/whisperkit" \
|
|
--max-workers 4
|
|
```
|
|
|
|
Expected folder layout:
|
|
|
|
```text
|
|
~/Library/Application Support/Blitztext/models/whisperkit/
|
|
openai_whisper-small_216MB/
|
|
AudioEncoder.mlmodelc/
|
|
MelSpectrogram.mlmodelc/
|
|
TextDecoder.mlmodelc/
|
|
```
|
|
|
|
If the folder is nested differently, the app will not detect the model.
|
|
|
|
## Notes
|
|
|
|
- First use can be slower because the model has to load and prewarm.
|
|
- Local transcription avoids sending audio to OpenAI for the Blitztext workflow.
|
|
- The app currently supports local transcription only, not local rewriting.
|
|
- Models are downloaded on demand so the repository and app package stay small and auditable.
|