Zerm latest — dictation and read aloud, running on-device. Release notes
Zerm
GitHub Download

Three on-device models · macOS

Dictate anywhere.
Read anything aloud.

Press a shortcut and Zerm drops clean text at your cursor. Select any text and it reads back in a natural voice. Parakeet, Whisper, Kokoro, and Gemma run locally — fast, private, offline. Cloud providers stay optional, and off by default.

Download for macOS checking latest… View source GPLv3 · Swift
  • macOS 14.4+
  • Apple Silicon
  • No account
  • GPLv3

Local runtimes bundled and managed for you

The engine room

Three models. All on your Mac.

Zerm downloads and manages each one for you. No API key, no signup, no round trip — the whole loop from microphone to pasted text to spoken playback can run with the network off.

01

Parakeet & Whisper

Speech → Text

Dictation and file transcription. Parakeet runs on FluidAudio and Whisper on whisper.cpp, including ivrit.ai models tuned for Hebrew. Zerm recommends the best fit for your Mac.

FluidAudio · whisper.cpp
02

Kokoro

Text → Speech

The Read Aloud voice. Runs on sherpa-onnx and sounds natural enough to listen to a long document without fatigue.

sherpa-onnx
03

Gemma

The agentic layer

Cleans up dictated text and rewrites messy source into something worth hearing. Runs on llama.cpp.

llama.cpp

Dictation

Talk. It types.

A global shortcut works in every app — editor, terminal, browser, Slack. Zerm handles the parts that usually break: permissions, the active insertion point, and pasting without stealing focus.

  1. Hold the shortcut Push-to-talk, or tap to toggle with auto-stop on silence.
  2. Speak normally Zerm transcribes locally while you are still talking.
  3. Release Clean text is pasted at your cursor, or copied if you prefer.
Native writing layer verification

Read aloud

Select text. Hear it.

Zerm reads whatever you have highlighted, anywhere in macOS. Before it speaks, Smart Reading cleans the text so the voice does not read your markup out loud.

  • Acronyms and units expanded the way a person would say them
  • URLs, code blocks, and tables summarised rather than spelled out
  • Emoji and markdown syntax dropped
  • Optional rewrite pass by the on-device LLM

Everything else

The rest of the workflow.

Power Mode

Detects the active app and browser URL, then applies that app's model, language, output, or prompt — without touching your global settings.

Personal dictionary

Teach it the names, products, and jargon that transcription models reliably get wrong.

History & Dashboard

Every transcript kept locally, searchable, copyable, and deletable. The Dashboard shows words and time saved for the range you pick.

Transcribe File

Drop in audio or video and get a transcript with each speaker identified. Rename speakers, then export TXT, Markdown, SRT, WebVTT, or JSON.

Local or cloud

Parakeet, Whisper, or Apple for speech; Gemma or Ollama for the LLM; plus eleven cloud speech providers and any AI provider you configure yourself.

Honest permissions

Microphone, Accessibility, and screen context are each requested in context — and setup failures are surfaced, not swallowed.

Privacy model

The boundaries, stated plainly.

Zerm is a voice app with microphone and Accessibility access. That deserves specifics rather than a slogan.

  • Local by default. Audio stays on the machine on the local transcription path.
  • Cloud is opt-in. No provider is contacted until you add a key and select it.
  • Request storage is separate. Logging AI payloads is its own setting, off from runtime calls.
  • Power Mode context is scoped. Active app and URL are read only for workflows you configure.
  • Auto-paste needs Accessibility. Decline it and Zerm falls back to the clipboard.
  • Auditable. All of the above is in the source, under GPLv3.

Download

Install it, or build it yourself.

Links resolve straight to the assets of the latest published release — currently loading….

macOS · Apple Silicon .dmg M1 and newer macOS · archive .zip Unzip into /Applications Source xcodebuild Xcode 16 · Swift & SwiftUI

Origin & license

Built on VoiceInk. Credited in public.

Zerm is a modified GPLv3 derivative of VoiceInk by Beingpax, which provided the foundation for the macOS app architecture, dictation workflow, transcription pipeline, Power Mode, and model management. The license stays, the attribution stays, and upstream is linked from everywhere it should be.

Upstream Beingpax/VoiceInk The original GPLv3 project Attribution NOTICE Migration and modification notes License GNU GPLv3 Source license for Zerm

Read the code. File issues. Send patches.

The repository carries the architecture notes, build instructions, and contributor guidance.

Open repository github.com/arcusis/Zerm Issues Bugs & requests