ParrotFlow vs Superwhisper

What Superwhisper does well

You pick the speech model for each mode. Local options include Whisper models, Parakeet and Cohere Transcribe. Cloud options include Superwhisper’s own S1-Voice, Deepgram Nova and ElevenLabs Scribe. With S1-Voice or a multilingual Whisper model, it covers more than 100 languages.

One Pro license covers all your Macs and Windows PCs, plus your iPhone and Android phone. Pro also transcribes audio and video files. On a Mac, Meetings records a call with no bot in it. It labels each speaker and writes a summary. Meetings arrived in version 3.0, in October 2026.

Superwhisper also connects to coding agents on the Mac, such as Claude Code, Codex and Cursor. When the agent asks a question, Superwhisper opens and you answer by voice.

The free plan keeps your dictation on your device. It gives unlimited dictation with local Whisper models, a vocabulary, replacements and up to two modes without AI. New users get 3,000 words of Pro to try.

Modes and steps

In Superwhisper, a Custom mode has a Custom instructions field. You describe the change in plain language, and the AI applies it to every dictation in that mode. Message mode removes filler words and fixes grammar, with instructions Superwhisper wrote. AI modes need Pro. By default they run on S1-Language, Superwhisper’s own cloud model. You can also pick a model from Anthropic, OpenAI, Google or Groq, or use your own API key. For dictation, the docs list one local language model, S1-mini. It is English only and cannot run Custom modes.

Superwhisper also has replacements, which use no AI. Its docs describe them as From → To pairs, matched regardless of case. They do not mention patterns.

In ParrotFlow, each step is a replace: table of patterns, a command: script, or a prompt: to a model you name. Rules and scripts need no model. A script is any program that reads stdin and writes stdout. A prompt runs through Ollama on your Mac, or on the OpenAI or Anthropic API with your own key.

One regex rule turns every spoken PR number into a link. A From → To table needs one row per number.

Some rules need code. The shipped slack_mentions script looks names up in a roster you fill, so it cannot invent a handle.

Any step can carry app: /slack/, so it runs in Slack and nowhere else. See programmable dictation, your first rule and a Python script on your dictation.

When the model runs

In Superwhisper, the mode you dictate in decides what the AI does. You switch modes with a shortcut, from the menu bar, or by app or website. To change a dictation afterwards, you open it in History and reprocess it with another mode.

In ParrotFlow, a prompt step can wait until after you speak. This one puts a Bug report button on the pill:

transforms:
  - name: Bug report
    prompt: Turn this into a bug report.
    offer: true
    key: b

Dictate a rough note and press B. The note becomes a bug report. If you do nothing, your words stay as you said them. The bug report guide has a fuller prompt.

Names

Superwhisper sends your vocabulary words to the voice model as hints. Its docs warn that too many words can confuse the model. In Super mode, a language model reads the text field and title of the active window, then corrects names.

When you correct a name, ParrotFlow offers to learn it. Press Y to keep it. Next time, it matches the name by spelling or by sound, then checks that the name fits the sentence. Word lists and a small sentence model do the check, not a large language model. It can still get a name wrong. Correct it back, and that counts against that use.

ParrotFlow also spells names the way your screen writes them. When you press the key, it reads the window you dictate into. Since version 0.18.0, it does this in any app. It skips password managers, System Settings and password fields. If it cannot read the window, it uses the text above the caret. This happens on your Mac.

See custom vocabulary and screen context.

Also compared: Wispr Flow and VoiceInk. Install with brew install znat/tap/parrotflow, or see Download.