Setup
Models
Which transcription model to use, what each one costs you in speed, size and memory, and how local and cloud differ.
Zerm separates the model that turns speech into text from the model that improves that
text afterwards. This page covers the first one. For the second, see
AI enhancement.
The same model list serves dictation and Transcribe File.
Finding a model
The models screen has four filters:
- Recommended — the default view. Zerm looks at your chip, memory, and macOS version
and picks the best fit for English, for many languages, and for Hebrew, plus the
flagship model of every cloud provider you have added a key for. - Local — every on-device model. Narrow it with English only, Multilingual,
or Great in Hebrew. - Cloud — every cloud model, filterable by provider.
- Custom — OpenAI-compatible endpoints you add yourself.
Cards carry badges for what matters when choosing: Streaming, Speaker labels,
Uses Dictionary, and how well the model handles Hebrew (Great in Hebrew,
Hebrew, or No Hebrew).
Local models
Local models download once and then run entirely on your Mac. No key, no account, no
network — dictation keeps working on a plane.
| Model family | Size | Languages | Notes |
|---|---|---|---|
| Sherpa-ONNX Moonshine Tiny EN | 103 MB | English only | Useful Sensors model converted to ONNX. |
| Sherpa-ONNX Moonshine Base EN | 239 MB | English only | Useful Sensors model converted to ONNX. |
| Sherpa-ONNX Zipformer Russian | 58 MB | Russian | Vosk model converted to Zipformer ONNX. |
| Parakeet Unified | 614 MB | English only | Best English model for most Apple Silicon Macs. |
| Parakeet 110M | 228 MB | English only | Small and fast. |
| Parakeet TDT v2 | 1.2 GB | English only | NVIDIA model through FluidAudio. |
| Parakeet Ultra / V3 | 595 / 494 MB | English + 25 European languages | Support real-time streaming. No Hebrew. |
| Parakeet Redux | 220 MB | English + 25 European languages | Compact streaming model. Requires macOS 15 or later. No Hebrew. |
| Whisper Tiny, Base, Small, Medium, Large v2/v3, and Large v3 Turbo | 75 MB–3.1 GB | Multilingual | Official ggml-org conversions of OpenAI Whisper. English-only variants available for Tiny, Base, Small, and Medium. |
| Whisper quantized variants | 44 MB–1.7 GB | Matches base model | ggml-org Q5_0 and Q8_0 files where published. Smaller files use less storage and may trade some accuracy. |
| Distil-Whisper Large v3 | 1.5 GB | English only | Official ggml release for whisper.cpp. |
| ivrit.ai Large v3 Turbo / Large v3 | 1.6 / 3.1 GB | Hebrew and English | Whisper tuned for Hebrew by ivrit.ai. Large v3 needs a Mac with 24 GB of memory or more. |
| Apple Speech | — | Multilingual | Built into macOS. Nothing to download. Requires macOS 26. |
Earlier versions retired Whisper Tiny, Base, and Small variants and Parakeet TDT v2. Those
models are offered again; the retired-model cleanup migration is disabled and preserves
existing selections and files.
Parakeet runs through FluidAudio and needs Apple Silicon. Whisper models run on
whisper.cpp, and you can import your own fine-tuned one — see
custom local models.
The ivrit.ai models treat Auto as Hebrew and keep the English words you mix in.
Choose English only when you will dictate nothing but English.
Apple Speech offers Hebrew only if your copy of macOS reports that it supports it.
Sources and licenses
Each downloadable local model card includes its creator, official model card and download
source, license, required credit, and conversion or quantization credit. Whisper downloads
are pinned to official ggml-org or ivrit.ai files and verified against SHA-256. Parakeet
models use FluidAudio's official Core ML model packages. Sherpa-ONNX model cards link to the
upstream model page and list sherpa-onnx as the ONNX converter.
Before the first download, Zerm shows a third-party model notice with creator, source, license,
and attribution. Downloads show progress and support pause, resume, cancel, and delete. Choose
Model credits to review attribution and open source and license links for installed models.
Streaming and live preview
With streaming, text appears in the recorder while you are still talking.
- Parakeet Ultra, Parakeet Redux (macOS 15+), and Parakeet V3 are the local models that stream. Their cards have
a Real-time switch, on by default once the model is downloaded. With voice activity
detection on, the live preview skips its passes while you pause. - Cloud models marked Streaming get the same Real-time switch once their key
is configured. - Show Live Text Preview, in Model Settings, decides whether the recorder shows that
text as it arrives. On by default.
When you stop, Zerm transcribes the complete recording once more, so the last words you
said are never cut off by the live preview.
Cloud models
Cloud models send your audio to a provider. Nothing is contacted until you add a key and
select the model, and you can keep local and cloud models installed side by side.
| Provider | Models |
|---|---|
| OpenAI | GPT Transcribe |
| Gemini | Gemini 3.5 Transcribe |
| ElevenLabs | Scribe V2 |
| Soniox | Soniox V5 |
| AssemblyAI | Universal 3.5 Pro, Universal 2 |
| Deepgram | Nova 3, Nova 3 Medical |
| Groq | Whisper Large v3 Turbo, Whisper Large v3 |
| Mistral | Voxtral Mini Transcribe 2 |
| Speechmatics | Speechmatics |
| Gladia | Solaria |
| xAI | Grok |
ElevenLabs, Soniox, Deepgram, Speechmatics, and xAI support streaming.
Open Configure on a card to add and verify the key. Use cloud models when you need a
language the local models handle poorly, or accuracy beyond what your Mac can run.
Custom models
If your provider exposes an OpenAI-compatible transcription API, add it under the
Custom filter with Add Model: a display name, endpoint, key, and model name.
Presets fill in the details for Together AI, DeepInfra, and OpenRouter.
Zerm sends a one-second test recording before saving. If the endpoint does not answer
correctly, nothing is saved and the error is shown in the form. Providers with their own
bespoke API are not supported through this path.
Model Settings
The Settings button on the models screen opens the settings for the selected model.
It shows only what that model actually supports, so the list changes as you switch.
- Output Format — a style hint the model follows, kept separately for each language.
Shown for models that accept a prompt, such as Whisper, OpenAI, Groq, Soniox, and
custom endpoints. - Automatic text formatting — paragraphs and capitalisation for long dictation. On by
default. - Voice Activity Detection and Prewarm model — for Whisper and Parakeet.
- Show Live Text Preview — while a streaming model is active.
- Cloud timeout — how long to wait for a cloud provider, from 30 seconds to 30
minutes. The default is 60 seconds.
Your dictionary terms are sent automatically to every model marked
Uses Dictionary. There is nothing to switch on.
Choosing a model
Disk and memory. Each card shows the download size. Zerm marks a model Heavy for
this Mac when it would use a large share of your memory, and Unavailable when it
would not fit.
Language. Set the language on the models screen. Parakeet and Gemini detect it
themselves, and English-only models are always English. When the model you pick does not
support your language (Parakeet, Voxtral and Grok have no Hebrew, for example), the
language settings say so, and Zerm lets the model detect the language instead of asking it
for one it does not know. A Power Mode can use a different model and
language for one app.
My languages. With a Whisper model and Auto-detect, add the languages you actually
speak under My languages. Detection then chooses only among them, so a short Hebrew
or Russian phrase cannot come out as a neighbouring language such as Arabic. When two of
your languages are about equally likely, the one of your current keyboard wins. Leave it
empty to allow every language.
Intel Macs. Local models do not run reliably on Intel hardware and Zerm says so.
Recommended picks only Whisper models there; a cloud model is usually the better choice.
Downloading and managing
Models download in the background with progress shown, and are verified after download.
Delete one at any time to reclaim the space. Deleting the model you are using clears your
default, so pick another before you dictate again.
Prewarm model is on by default: shortly after launch and after your Mac wakes, the
active Whisper or Parakeet model is loaded before you press anything, so the first
dictation is not the slow one. It is skipped in Low Power Mode.
The enhancement model is separate
If you use enhancement, that runs on a second model — on-device via
llama.cpp, Ollama, a local CLI, or a cloud provider. It is chosen and downloaded
separately.
On device, enhancement offers Gemma 4 E2B (the default) and Qwen3 4B Instruct.
Read Aloud picks its on-device model on its own screen.