Wispr Flow, minus the cloud and the subscription

What it does

BetterWispr is hold-to-talk dictation for macOS: hold a shortcut, speak, release, and the text is pasted into whatever app had focus. Transcription happens on-device via Apple’s speech recognition, NVIDIA Parakeet, or Whisper large-v3 turbo. If paste can’t safely reach the previous app, it falls back to the clipboard so a dictation is never lost.

The interesting bit

It learns from your corrections — fix a misheard name once and it’s spelled right next time — and supports voice commands like “scratch that” and “at the rate KV” for @KV. There’s also a meetings mode with live transcription of both sides of a call, plus on-device summaries and action items via Apple Intelligence or a local Ollama model.

Key highlights

• Fully local by default; audio never leaves your Mac unless you explicitly pick a cloud option
• Multiple engines: Apple speech (no download), Parakeet v2/v3, Whisper large-v3 turbo
• Vocabulary hints and phrase replacements for names and jargon
• Transcribe audio/video files or links (via yt-dlp) to timestamped Markdown
• Optional MCP access for Claude, Cursor and friends to read your transcripts — off by default, read-only
• Bring your own endpoint: Sarvam AI, Smallest AI, or any OpenAI-compatible URL; keys live in the Keychain

Caveats

• macOS 14+, Apple Silicon recommended; Intel Macs are second-class here
• Settings, vocabulary and history sit in an unencrypted workspace.json
• The optional Codex subscription mode uses an unofficial ChatGPT endpoint — fine, but know what you’re opting into
• First transcription with a large model is slow to load

Verdict

If you dictate a lot and balk at sending your voice to someone else’s servers, this is one of the more complete local-first options on macOS. If you’re not on a Mac or don’t care about cloud dictation, move along.