Free SuperWhisper Alternatives: Local Voice Typing

Short answer: install Handy first. It is MIT-licensed, runs fully offline on macOS, Windows and Linux, and had 29,837 GitHub stars on August 18, 2026 — created February 2025, so roughly 30,000 stars in eighteen months. On Linux, budget an extra twenty minutes: Handy's own README lists Wayland support as limited and needs wtype, dotool or (on Ubuntu 26.04) ydotool before text injection works reliably.

If you are on Linux full-time, skip Handy and go to hyprwhspr or Vocalinux instead — both are built Linux-first and both state Wayland and X11 support in their own documentation. If you are on macOS and want something that feels native, OpenSuperWhisper is a 2,634-star Swift app installable with brew install opensuperwhisper.

One correction before we go further, because it changes what you are actually shopping for.

---

What You Are Actually Escaping

Neither SuperWhisper nor Wispr Flow is a pure paywall — both have free tiers. What you cannot buy at any price from either is a Linux build.

Here is what the two incumbents' own pricing pages said on August 18, 2026:

So the honest reasons to move to an open-source app are, in order: you are on Linux and neither product exists for you; you hit the word cap and $15/month for dictation feels absurd; or you want the audio to never leave the machine, which a weekly word quota strongly implies is not the case on the free cloud tiers.

If none of those is you, the free tier of the thing you already have is fine. We would rather tell you that than sell you a migration.

---

The Comparison Table

Star counts, licences and last-push dates below were read from the GitHub API on August 18, 2026. Everything in the "backend" and "text injection" columns comes from each project's own README.

One warning on that last row. WhisperWriter is still the app most often recommended in older forum threads, but its last push was August 24, 2024 — two years stale as of this writing. It may still work; it is not being maintained at the pace of the rest of this list, and we would not start there in 2026.

Note the licence column if this is for work. Handy, OpenWhispr, OpenSuperWhisper and hyprwhspr are MIT. Both TypeWhisper builds and WhisperWriter are GPL-3.0. Vocalinux is AGPL-3.0, which is the most restrictive of the three if you ever plan to build a service around it.

---

Mac

On macOS you have the most choice and the least friction: OpenSuperWhisper for something small and native, TypeWhisper for something feature-heavy, Handy if you also use Windows or Linux and want one app everywhere.

OpenSuperWhisper is a Swift app, MIT, Apple Silicon only (Intel support is on its contribution TODO list, not shipped). Install is one line — brew install opensuperwhisper — and it offers two engines you download from inside the app: Whisper and Parakeet. Its trigger options are the most flexible of the bunch: a key combination, a single modifier key such as Left ⌘ or Right ⌥, or a mouse button (middle or a thumb button), each with hold-to-record.

TypeWhisper for Mac is the opposite philosophy — eleven transcription engines including WhisperKit, Parakeet TDT v3, Apple SpeechAnalyzer, Granite Speech, Qwen3 ASR and local Voxtral Mini 4B, plus workflows, snippets and a plugin SDK. It requires macOS 14 (Sonoma) or later, and Microphone plus Accessibility permission. It is GPL-3.0 with a commercial licence offered separately, and there is a supporter/premium tier, so read its licence page before you deploy it across a team.

A macOS-specific gotcha from Handy's README that applies to any of these: shortcuts built on the fn (Globe) key only fire on Apple keyboards. Third-party keyboards handle Fn in firmware and send nothing to the Mac, so there is no event for the app to catch. If you switch between the MacBook keyboard and an external one, pick a shortcut made of standard modifiers.

If you are building out the rest of a local stack on the same machine, our Mac local AI setup guide covers the neighbours.

---

Windows

Handy and OpenWhispr both ship Windows builds; TypeWhisper for Windows is the Windows-native option and explicitly states its recommended local models run on CPU with no GPU required.

That last point is the practical one. TypeWhisper Windows supports Windows 10 or later on x64 and ARM64, and lists on-device models running through plugins: Parakeet TDT 0.6B (around 700MB of disk), Canary 180M Flash (around 200MB), whisper.cpp models and Granite Speech. Its own recommendation table pairs "fast general dictation" with Parakeet TDT 0.6B or whisper.cpp Base Q5_0, and "lowest disk usage" with whisper.cpp Tiny Q5_0. For an ARM laptop with no discrete GPU, that is the whole reason this category became usable.

Handy's README lists a real, unresolved issue worth knowing before you commit: Whisper models crash on certain Windows and Linux configurations, marked configuration-dependent and flagged help-wanted. If you hit it, switching to the Parakeet backend sidesteps the Whisper code path entirely.

For the backend question underneath all of this — when Parakeet beats Whisper and when it does not — see our Parakeet vs Whisper comparison.

---

Linux (Read This First)

Linux is where these apps fail, and the failure is almost never the speech model. It is text injection under Wayland.

Under X11, xdotool synthesises key events into any window and everything works. Wayland deliberately removed that capability for security reasons, so each app has to pick a workaround, and the workarounds have different coverage:

• wtype — the common choice, works on compositors implementing the virtual-keyboard protocol. Handy calls it the preferred Wayland tool.
• ydotool — works at the kernel uinput layer, so it works everywhere, but needs a privileged daemon and a systemd unit. Handy's README states that on Ubuntu 26.04, which defaults to Wayland, wtype does not work and you need ydotool configured with systemd.
• dotool — similar approach, requires adding your user to the input group (sudo usermod -aG input $USER, then log out and back in).
• IBus — the input-method route Vocalinux uses on Wayland, with specific handling documented for KDE Plasma Wayland when ibus-wayland is absent.

The second Wayland problem is global hotkeys. Under Wayland the compositor owns keyboard shortcuts, so an app cannot simply grab one. Handy's answer is to expose CLI flags and let you bind them in your desktop environment:

# bind these in GNOME/KDE custom shortcuts, or in your WM config
handy --toggle-transcription
handy --toggle-post-process

# Hyprland (~/.config/hypr/hyprland.conf)
bind = $mainMod, O, exec, handy --toggle-transcription

# Sway / i3
bindsym $mod+o exec handy --toggle-transcription

If Linux is your only OS, use a Linux-first app. hyprwhspr installs from the AUR (yay -S hyprwhspr) or a shell installer on Debian/Ubuntu/Fedora/openSUSE, states Wayland and X11 support, and requires wl-clipboard and wtype on Wayland or xclip/xdotool/xprop on X11. It offers Parakeet TDT V3, Whisper and onnx-asr locally, with cloud APIs strictly optional. Vocalinux is AGPL-3.0, defaults to whisper.cpp with Vulkan GPU acceleration — which means AMD and Intel GPUs work, not just NVIDIA — and falls back to VOSK on low-RAM machines. Its default model download is about 74MB.

Two more Handy-on-Linux specifics that will otherwise cost you an afternoon:

1. Startup crash with libgtk-layer-shell.so.0 — install libgtk-layer-shell0 (Ubuntu/Debian), gtk-layer-shell (Fedora/Arch). This is the most common startup failure in its own troubleshooting section.
2. The recording overlay steals focus on some compositors, so the paste lands in the wrong window. Settings → Advanced → Overlay Position → None. It is disabled by default on Linux for exactly this reason.

---

Latency, Honestly

We have not benchmarked key-release-to-text latency on these apps ourselves, so we are not going to print a table of numbers we did not measure. Anyone who publishes one without naming the machine, the model and the display server is guessing.

What we can tell you is what the number is made of, so you can predict yours:

1. Model and backend. A Whisper tiny or base quant returns in a fraction of the time a large-v3 model does. Handy's README cites Parakeet V3 at roughly 5x real-time on mid-range hardware, tested on an i5, in CPU-only operation — so a five-second utterance is about a second of compute on that class of machine, before any of the other costs below.
2. Whether it streams. TypeWhisper for Mac advertises a streaming preview with WhisperKit, showing partial text while you speak. That does not make the final result faster, but it removes the dead-air feeling that makes a two-second wait feel like four.
3. Whether the model stays resident. hyprwhspr's pitch is in-memory models for near-instant response, with explicit unload/reload commands to free VRAM. An app that loads the model per utterance pays the load cost every time.
4. The injection path. Clipboard-and-paste adds a round trip and, on Linux, a focus race. Direct key synthesis is faster and more reliable.

If a specific app feels laggy, change one of those four things — usually the model — before you switch apps. Our faster-whisper guide covers the speed levers on the Whisper side, and the WhisperX guide covers the batched pipeline if you also do bulk transcription.

---

How to Choose

Four questions, in this order:

1. What OS? Linux-only → hyprwhspr or Vocalinux. macOS-only → OpenSuperWhisper (light) or TypeWhisper (heavy). Windows-only → TypeWhisper for Windows or Handy. More than one → Handy, because it is the only one on this list actively shipping all three.
2. Do you have a GPU? If not, choose an app that defaults to Parakeet or a small whisper.cpp quant. Both TypeWhisper Windows and Handy document CPU-first paths.
3. Wayland or X11? If Wayland, check the app's injection tool before installing, and expect to install wtype or ydotool.
4. Does the licence matter? MIT for Handy, OpenWhispr, OpenSuperWhisper and hyprwhspr. GPL-3.0 for both TypeWhispers. AGPL-3.0 for Vocalinux.

And the thing worth saying plainly: dictation is a different job from transcription. If what you actually want is to turn recordings or videos into text, none of these is the right tool — see Whisper local speech-to-text for files and local AI subtitles for video. If you want to talk to a model rather than type with your voice, that is a local voice assistant.

---

What We Could Not Verify

We read every one of these projects' documentation and confirmed the repository metadata directly against the GitHub API. We did not install and stopwatch each app on macOS, Windows, X11 and Wayland — that is four operating environments and eight apps, and publishing invented latency figures would be worse than publishing none. Specifically unverified by us, and taken from project documentation where cited: per-app latency, idle RAM, and whether each app's injection reaches every native application on every desktop. Where a project's own README flags a limitation — Handy's Whisper crashes on some Windows and Linux configurations, its limited Wayland support, OpenSuperWhisper's Apple-Silicon-only requirement — we have passed that on rather than smoothing it over.

Star counts move. The ones here are a snapshot from August 18, 2026.

---

Sources

• GitHub REST API, repository metadata for each project (stars, licence, creation and last-push dates), queried August 18, 2026
• Project READMEs: cjpais/Handy, OpenWhispr/openwhispr, Starmel/OpenSuperWhisper, TypeWhisper/typewhisper-mac, TypeWhisper/typewhisper-win, goodroot/hyprwhspr, VocaHQ/vocalinux, savbell/whisper-writer
• superwhisper.com pricing table and wisprflow.ai/pricing, as displayed August 18, 2026

---

FAQ