Manage models

TranscribeX runs speech models on your Mac. You download the ones you want, and you can delete them at any time to free up disk space. All model settings live under Settings › Transcription Models. Open Settings from the gear icon at the top right of the main window.

Not sure which model to pick? See choosing a model.

The model panes

Transcription Models has one pane per model family:

• Whisper Model: OpenAI Whisper models, including distilled and language-tuned versions.
• Parakeet Model: NVIDIA Parakeet and Nemotron models, including the streaming models used for live transcription.
• Qwen Model: Qwen3 ASR models.
• Cohere Model: Cohere Transcribe models.
• Mistral Online and ElevenLabs Online: cloud models that need your own API key (Pro).

Each local pane has a Model List with the model name, its download size and a short description. Models marked Recommended are a good starting point. Some Qwen and Cohere rows say Timestamps not supported: those models give you text without segment timing.

Download a model

1. Open Settings › Transcription Models and choose a pane.
2. Click the download arrow next to the model. A progress bar shows while it downloads.
3. When it finishes, the row shows a folder icon and a trash icon.

Free models (Whisper Tiny, Parakeet EOU 160ms and Apple On-Device) download for everyone. Other models show a Pro button instead of the download arrow until you activate Pro.

Note: The first time you use a model, loading it can take a few minutes. Later runs start much faster.

Need a fine-tuned Whisper model from Hugging Face that isn’t listed? Email support@transcribex.io.

Delete a model

Click the trash icon next to a downloaded model. Click the folder icon to see its files in Finder. If TranscribeX can’t remove the files, it shows Delete Failed, and you can delete the folder yourself in Finder.

Change where models are stored

The Whisper Model, Qwen Model and Cohere Model panes each have a Model Storage Path. Use it to keep large models on another drive.

1. Delete every downloaded model in that pane first. TranscribeX won’t change the path while models are still there and shows Cannot Change Model Storage Path with a list of the models to remove.
2. Click … next to Model Storage Path and choose a folder.
3. Download the models again. They go to the new folder.

To go back to the default location, delete the models in that pane, then click Reset.

Compute units

Compute units decide which part of your Mac runs the model. There are two settings:

• Whisper Model › Audio Encoder for Whisper models.
• Parakeet Model › Compute Unit for Parakeet and Nemotron models.

Both offer the same four options. The default is Neural Engine.

Distilled Whisper models always run on the GPU when you choose Neural Engine or All.

Temperature

The Whisper Model pane also has Temperature, from 0.0 to 1.0. It controls how random the output is. The default, 0.0, gives the most predictable results, and most people should leave it there. Reset to Defaults puts Audio Encoder and Temperature back to their defaults.

Model Updates Available

When a Whisper model you downloaded has an outdated configuration, the Whisper Model pane shows a Model Updates Available banner listing the models to update, and those rows get a warning icon. To update one, delete it with the trash icon and download it again.

Set your default model

The default model is used for new files, recordings and bulk jobs. Set it in Settings › Transcription › Default Model. The menu lists the models you can use, with a green check next to the ones already downloaded. If you pick one that isn’t downloaded, the pane warns you that it “is selected as the default model but is not downloaded yet.” Apple On-Device appears in this menu on macOS 26 or later.

You can always pick a different model for one file from the model menu at the top of the transcript. More Models... at the bottom of that menu opens the model settings.

Use Mistral or ElevenLabs online models (Pro)

Online models send your audio to the provider’s servers, use your own API key, and are billed by that provider. Local models process audio entirely on your Mac.

1. Open Settings › Transcription Models › Mistral Online or ElevenLabs Online.
2. Turn on Enable Mistral or Enable ElevenLabs.
3. Paste your key into API Key.
4. Optionally turn on Diarize so the provider labels segments by speaker.

The provider’s models then appear in the model menu and in Default Model. Mistral offers voxtral-mini-2602 and voxtral-mini-latest. ElevenLabs offers scribe_v1 and scribe_v2 for files, plus Scribe v2 Realtime for live transcription. Online models don’t report progress while they work, so a long file can sit at the start for a few minutes.

Note: With Mistral Diarize on, Mistral doesn’t return word timestamps, so Reflow Transcripts won’t work on those transcripts.