# Install the local speech model

> Download, import, and troubleshoot the one-time Whisper model that powers on-device transcription.
>
> Verified against the current RoughCut app on 28 August 2026
> https://www.roughcuteditor.com/docs/speech-model

On-device transcription needs one file, downloaded once. RoughCut uses the open-source Whisper model ggml-large-v3-turbo.bin, about 1.6 GB, from the whisper.cpp project on Hugging Face. The normal path is Settings, then Models & Data, then Download from Hugging Face, which shows live percentage and byte progress. If that download stalls or fails, use Copy download link to fetch the exact file in a browser and Import model file to point RoughCut at it. The model is cached in the app container, used with Metal acceleration, and never downloaded again. You only need it for on-device transcription.

## Prerequisites

- About 2 GB of free disk space.
- A network connection for the initial download.
- Nothing else. The model is free and open source, and no account or key is involved.

You can skip this entirely if you have RoughCut Pro and intend to use [Cloud transcription](/docs/transcription-options) for every job — that path uses OpenAI instead of a local model.

## Steps

1. Open `Settings` and select the `Models & Data` tab.
2. Under the speech model section you will see either `Model ready (ggml-large-v3-turbo.bin)` or the notice `The speech model (~1.6 GB) is required for transcription.`
3. Click `Download from Hugging Face`.
4. Watch the progress. RoughCut shows `Downloading model… <n>%` and, once byte counts arrive, `Downloading model… <n>% (<written> of <total>)`. The same live status also appears on the `New Rough Cut` and `Transcribe Only` screens, so you do not have to sit in Settings.
5. When it finishes, the section switches to `Model ready` and `Process` becomes available.

## Manual install when the download will not finish

RoughCut exposes the exact file rather than hiding it, so you can always fall back to a browser.

1. In `Settings → Models & Data`, click `Copy download link`. RoughCut copies the address to your clipboard and confirms with `Model download link copied.`
2. Paste it into a browser and download the file. The address is:
   ```text
   https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin
   ```
3. Back in `Settings → Models & Data`, click `Import model file…`.
4. Select the `.bin` file you downloaded. RoughCut copies it into its own models folder, replacing any existing file of that name.

## Expected result

The model section reads `Model ready` with the model name, and transcription runs locally. RoughCut caches exactly one copy in its application container and never downloads it again.

## Where the model lives and how it runs

The file is stored inside RoughCut's own Application Support container, in a `models` folder. Because that lives in the app's sandbox, **deleting RoughCut removes the model**, and you would download it again on a fresh install.

Transcription runs with Metal acceleration on the GPU by default. The `Whisper device` setting offers `auto`, `cpu`, and `metal`; `metal` is correct on Apple Silicon and `cpu` exists for debugging and is much slower.

Advanced users can point RoughCut at a model stored elsewhere using the `Whisper model path` setting, which overrides the default location. Leave it empty unless you have a specific reason.

## Common mistakes

- **Downloading a different Whisper model.** RoughCut expects `ggml-large-v3-turbo.bin`. It is the tested default for the word-level timestamps that the cut plan and SRT export both depend on.
- **Importing a partial download.** RoughCut treats the model as ready only when the file is larger than roughly 100 MB, which catches truncated downloads — but a partly written file that clears that bar can still fail at transcription time. If transcription behaves strangely after a manual import, download the file again.
- **Expecting the model to survive app deletion.** It does not.
- **Waiting on the model when you do not need it.** Cloud transcription (Pro) needs no local model at all.

## Troubleshooting

| Symptom | Likely cause | What to do |
|---|---|---|
| Progress sits at 0% | Network or host issue | Cancel and retry; if it recurs, use the `Copy download link` and `Import model file…` path above. |
| Download fails repeatedly | Corporate network or DNS filtering | Download in a browser, then import the file. |
| `Whisper model missing at <path>. Download or import it in Settings.` | The file is absent or the custom path is wrong | Check `Whisper model path` is empty, then download or import again. |
| Not enough disk space | Less than about 2 GB free | Free space, then retry. A failed write leaves an unusable file — download again rather than importing the partial one. |
| Transcription produces nothing | Wrong or corrupt model file | Re-download the exact filename above and import it again. |

If the job itself fails rather than the download, start from [Troubleshooting by symptom](/docs/troubleshooting).