goodbye Wisprflow
I saw this Wispr Flow alternative called Ghost Pepper. I thought it was good, but it doesn't come close to Wispr Flow. So I built my own on top of it, and for how I use it, it's better than Wispr.
It's called Diction. It's a dictation app for macOS. Hold a shortcut, talk, let go, and the text gets typed wherever your cursor is, in any app.
It costs about 5–10 cents a month. Everything runs on your Mac. The only paid part is a small Claude Haiku call that cleans up each transcript, and you use your own Anthropic API key. For me that comes to 5–10 cents a month. Wispr Flow Pro is $15 a month, so that's roughly 1/200 of the price.
A free, fully local mode is coming. I'm working on a mode that does the cleanup on your Mac too, with no API key and no cost. It's slower than Haiku, but it's free and nothing leaves your machine.
Features
Dictation
- Hold to talk, in any app. The text is pasted where your cursor is. The shortcut can be changed.
- Hands-free mode: tap a shortcut and keep talking without holding anything. Esc cancels.
- A small pill under the menu bar moves with your voice, so you can see it's listening.
- It pauses music and video while you dictate.
- "Paste last transcript" in the menu bar, plus a microphone picker.
Speech recognition, on your Mac
- Parakeet v3 by default. Whisper, Qwen3-ASR, Nemotron streaming and Apple's SpeechAnalyzer are also available.
- Your audio never leaves your Mac.
- Languages: automatic detection, English, German, Spanish, French, Italian.
AI cleanup (Claude Haiku)
- Removes filler words and fixes punctuation. Three levels: Off (exactly what you said), Light, and Medium (also tightens long sentences).
- Fast: it sends the request when you pause, before you let go of the key, and keeps the connection warm.
- If you're offline or the API is slow, a small local model takes over automatically.
- Very short dictations of 3 words or fewer skip the cleanup, so they're instant.
It gets your words right
- Context awareness: it can read the window you're dictating into, so names on screen are spelled correctly.
- Dictionary: add names and jargon, or fixes like "clawed → Claude".
- Learns from your corrections: if you fix a word after it's pasted, it adds that fix to your dictionary.
Swaps (voice snippets)
- Say a short phrase and get something longer typed instead, like your address, a link or an email you send often.
Styles
- Formal, Casual, lowercase or Excited.
History
- Everything you've dictated, searchable and grouped by day. Click a line to copy it.
- Your whole history drawn as one spiral line, with your total word count.
Meeting notes
- Records your mic and the call audio, and labels them "Me" and "Others".
- Transcribed on your Mac and saved as Markdown files, one folder per day.
- Can offer to start notes when a Zoom, Teams, Google Meet, FaceTime, Webex, Slack or Discord call begins.
Other
- Your API key is stored in the macOS Keychain.
- Open at login, optional Dock icon, sounds can be turned off.
- Open source (MIT).
Privacy
Audio stays on your Mac. With an API key, only the transcript text goes to Anthropic, plus the text it read from your window if context awareness is on. You can turn that off. The local mode will keep everything on your Mac.
Pricing
- Speech recognition: free, it runs on your Mac.
- Cleanup with Claude Haiku: your own Anthropic API key, about half a cent per 1,000 words (about 1 cent if it also reads your screen for names). At 1,000 words a day that's 10–15 cents a month. I pay 5–10 cents.
- Local only: $0.
- Wispr Flow Pro: $15 a month ($12 billed yearly). That's about 200× what I pay.
Credit
Diction started as a fork of Ghost Pepper by Matt Hartman (MIT). Thanks for the head start, Matt.
On first launch it downloads the speech model and a small cleanup model (~0.5 GB for the cleanup model). After that it works offline.
GitHub: bendrape1-byte/Diction
Happy to answer questions, and I'd love feedback, especially from people who use Wispr Flow every day.
87 points · 67 comments
Mentions
goodbye Wisprflow I saw this Wispr Flow alternative called Ghost Pepper. I thought it was good, but it doesn't come close to Wispr Flow. So I built my own on top of it, and for how I use it, it's better than Wispr. It's called Diction. It's a dictation app for macOS. Hold a shortcut, talk, let go, and the text gets typed wherever your cursor is, in any app. It costs about 5–10 cents a month. Everything runs on your Mac. The only paid part is a small Claude Haiku call that cleans up each transcript, and you use your own Anthropic API key. For me that comes to 5–10 cents a month. Wispr Flow Pro is $15 a month, so that's roughly 1/200 of the price. A free, fully local mode is coming. I'm working on a mode that does the cleanup on your Mac too, with no API key and no cost. It's slower than Haiku, but it's free and nothing leaves your machine. Features Dictation - Hold to talk, in any app. The text is pasted where your cursor is. The shortcut can be changed. - Hands-free mode: tap a shortcut and keep talking without holding anything. Esc cancels. - A small pill under the menu bar moves with your voice, so you can see it's listening. - It pauses music and video while you dictate. - "Paste last transcript" in the menu bar, plus a microphone picker. Speech recognition, on your Mac - Parakeet v3 by default. Whisper, Qwen3-ASR, Nemotron streaming and Apple's SpeechAnalyzer are also available. - Your audio never leaves your Mac. - Languages: automatic detection, English, German, Spanish, French, Italian. AI cleanup (Claude Haiku) - Removes filler words and fixes punctuation. Three levels: Off (exactly what you said), Light, and Medium (also tightens long sentences). - Fast: it sends the request when you pause, before you let go of the key, and keeps the connection warm. - If you're offline or the API is slow, a small local model takes over automatically. - Very short dictations of 3 words or fewer skip the cleanup, so they're instant. It gets your words right - Context awareness: it can read the window you're dictating into, so names on screen are spelled correctly. - Dictionary: add names and jargon, or fixes like "clawed → Claude". - Learns from your corrections: if you fix a word after it's pasted, it adds that fix to your dictionary. Swaps (voice snippets) - Say a short phrase and get something longer typed instead, like your address, a link or an email you send often. Styles - Formal, Casual, lowercase or Excited. History - Everything you've dictated, searchable and grouped by day. Click a line to copy it. - Your whole history drawn as one spiral line, with your total word count. Meeting notes - Records your mic and the call audio, and labels them "Me" and "Others". - Transcribed on your Mac and saved as Markdown files, one folder per day. - Can offer to start notes when a Zoom, Teams, Google Meet, FaceTime, Webex, Slack or Discord call begins. Other - Your API key is stored in the macOS Keychain. - Open at login, optional Dock icon, sounds can be turned off. - Open source (MIT). Privacy Audio stays on your Mac. With an API key, only the transcript text goes to Anthropic, plus the text it read from your window if context awareness is on. You can turn that off. The local mode will keep everything on your Mac. Pricing - Speech recognition: free, it runs on your Mac. - Cleanup with Claude Haiku: your own Anthropic API key, about half a cent per 1,000 words (about 1 cent if it also reads your screen for names). At 1,000 words a day that's 10–15 cents a month. I pay 5–10 cents. - Local only: $0. - Wispr Flow Pro: $15 a month ($12 billed yearly). That's about 200× what I pay. Credit Diction started as a fork of Ghost Pepper by Matt Hartman (MIT). Thanks for the head start, Matt. On first launch it downloads the speech model and a small cleanup model (~0.5 GB for the cleanup model). After that it works offline. GitHub: bendrape1-byte/Diction Happy to answer questions, and I'd love feedback, especially from people who use Wispr Flow every day.