Verdict: For anyone who wants accurate, private dictation without a subscription, Handy is the best free option in 2026. It runs entirely on your machine using OpenAI Whisper or NVIDIA Parakeet models, pastes text into any app via a keyboard shortcut, and costs nothing — no account, no cloud, no word limits. The trade-off is a 2–5 second processing delay after you stop speaking and occasional first-word clipping, but for privacy-conscious users, developers, and small-business owners, nothing else matches its combination of price (zero), privacy (100% local), and openness (MIT-licensed, 27,000+ GitHub stars).
Last verified: 2026-07-31 · Best free offline: Handy · Best paid: Wispr Flow ($15/mo) · Best Mac-only: Superwhisper ($9.99/mo) · Best built-in: Apple Dictation (free, Apple Silicon)
Pricing and model versions change often — last checked July 2026.
What Is Handy and How Does It Work?
Handy is a free, open-source, cross-platform desktop application that transcribes your speech entirely on your own computer and pastes the result into whatever text field is active. No audio is uploaded. No cloud servers are involved. You press a keyboard shortcut, speak, release, and your words appear.
The app was created by developer CJ Pais and is built as a Tauri application — a Rust backend with a React/TypeScript frontend. That makes it dramatically leaner than Electron-based alternatives; the Windows installer is roughly 12.6 MB. It runs on macOS, Windows, and Linux.
The transcription pipeline works like this:
- Press a configurable keyboard shortcut to start recording (toggle or push-to-talk mode).
- Silero VAD (Voice Activity Detection) filters out silence in real time so you're not wasting compute transcribing dead air.
- Release the shortcut. Handy sends the audio to your chosen model.
- The model transcribes locally and Handy pastes the text directly into whatever input has focus.
Which speech models does Handy support?
Handy lets you choose between two families of models, each with different speed-accuracy trade-offs:
| Model | Size | Runs on | Speed | Best for |
|---|---|---|---|---|
| Whisper Small | ~487 MB | CPU or GPU | Fast | Quick notes, decent accuracy |
| Whisper Medium | ~492 MB | CPU or GPU | Medium | Balanced quality |
| Whisper Large-v3 Turbo | ~1.6 GB | GPU recommended | 8× faster than Large-v3 | High accuracy with speed |
| Whisper Large-v3 | ~1.1 GB | GPU recommended | Baseline | Maximum accuracy |
| Parakeet V3 | ~478 MB | CPU only | ~5× real-time on mid-range i5 | Everyday dictation, low resource use |
Whisper Large-v3 achieves approximately 2.7% Word Error Rate (WER) on the LibriSpeech clean-audio benchmark and roughly 8–12% WER on real-world English audio. Parakeet V3, developed by NVIDIA as part of its NeMo framework, is CPU-optimized — it runs at about five times real-time speed on a mid-range Intel i5 chip, meaning a 10-second utterance transcribes in about 2 seconds.
You can also drop your own fine-tuned Whisper GGML models into Handy's models folder, restart the app, and they appear as custom options. That extensibility is central to Handy's design philosophy: as the project states, "Handy isn't trying to be the best speech-to-text app — it's trying to be the most forkable one."
Handy vs WhisperFlow: Is Free Actually Better?
The most common comparison is Handy versus WhisperFlow (also known as Wispr Flow), the leading commercial dictation app. Here's how they stack up:
| Feature | Handy (Free) | Wispr Flow Pro ($15/mo) |
|---|---|---|
| Price | Free, MIT-licensed | $15/month or $144/year |
| Offline | Yes, 100% local | No — cloud processing |
| Account required | No | Yes |
| Word limits | None | Free tier: 2,000 words/week |
| AI text cleanup | No (raw transcription) | Yes (filler removal, formatting, tone) |
| Processing delay | 2–5 seconds after release | Near real-time (cloud) |
| Custom models | Yes (drop in GGML files) | No |
| Privacy | Audio never leaves your machine | Audio sent to Wispr's servers |
| Platforms | macOS, Windows, Linux | macOS, Windows, iOS, Android |
| Language support | 99+ (Whisper), auto-detect (Parakeet) | 100+ |
| Open source | Yes (MIT) | No |
The accuracy gap is narrower than you'd expect. In real-world testing with technical jargon — terms like "SQLite," "GPU cluster," and "LLM" — both Handy (running Parakeet V3 or Whisper) and Wispr Flow correctly identified the majority of technical terms and added proper punctuation. Google Docs Voice Typing, by contrast, struggled badly with the same text, garbling "Grok" as "rock" and "ChatGPT" as "Chad GPT."
Where Wispr Flow pulls ahead is latency and post-processing. Because it processes in the cloud with dedicated infrastructure, transcription appears almost instantly, and its AI cleanup removes filler words, fixes capitalization, and can adjust tone. Handy gives you raw model output — accurate, but unpolished.
How Does Handy Compare to Other Free Dictation Options?
Handy isn't the only free dictation tool. Here's how it compares to the other major free and freemium options in 2026:
| Tool | Price | Offline | Open Source | Platforms | Key Limitation |
|---|---|---|---|---|---|
| Handy | Free | Yes | Yes (MIT) | Mac, Win, Linux | 2–5s delay, no AI cleanup |
| Apple Dictation | Free | Yes (Apple Silicon) | No | macOS, iOS | Apple devices only, no custom vocab |
| Windows Voice Access | Free | Partial (needs internet on most configs) | No | Windows 11 | English-dominant, limited languages |
| Google Docs Voice Typing | Free | No | No | Browser (Chrome) | Browser-only, no punctuation, cloud-only |
| OpenWhispr | Free tier (2,000 cloud words/week) | Yes (local mode) | Yes | Mac, Win, Linux | Free cloud tier capped; offline needs model download |
| Spokenly | Free (local models) | Yes | No | Mac, Win, Linux, iOS | Not open source; cloud features cost $9.99/mo |
Apple Dictation is the strongest free option for Mac users on Apple Silicon — it processes locally and is built into the OS. But it lacks custom vocabulary support and is locked to Apple devices. Google Docs Voice Typing is convenient but cloud-dependent and notably less accurate with technical terms. Handy's unique advantage is being fully cross-platform, fully open source, and completely free with no caps.
How to Set Up Handy in 5 Minutes
Getting started with Handy is straightforward. Here's the quick-start path:
Step 1: Download and install
Download the latest release from the Handy GitHub releases page or the project website. The latest version as of July 2026 is v0.9.x.
- macOS: Also available via Homebrew:
brew install --cask handy - Windows: Available via winget:
winget install cjpais.Handy - Linux: Download the AppImage or .deb package from the releases page
After installing, launch Handy and grant the necessary system permissions (microphone access and accessibility permissions for global keyboard shortcuts).
Step 2: Choose your model
Open Handy's settings and navigate to the Models section. Each model displays a speed and accuracy meter. For most users:
- Parakeet V3 is the best default — CPU-only, fast (~5× real-time on mid-range hardware), automatic language detection, and only ~478 MB to download.
- Whisper Small (~487 MB) if you want slightly different accuracy characteristics and have a GPU.
- Whisper Large-v3 Turbo (~1.6 GB) for maximum accuracy if you have a GPU and can tolerate a slightly longer download.
Step 3: Configure your shortcut
Set a custom keyboard shortcut in Settings. You can choose between:
- Toggle mode: Press once to start, press again to stop.
- Push-to-talk mode: Hold the key down while speaking, release when done.
Push-to-talk is generally more reliable because the VAD knows exactly when you're done speaking, which reduces first-word clipping.
Step 4: Start dictating
Press your shortcut, speak naturally, and release. Handy transcribes locally and pastes the text into whatever field is focused — your email client, code editor, Slack, a CMS, any text input. If transcription fails for any reason, Handy saves the audio recording so you can retry with a different model.
What Are the Known Limitations of Handy?
Handy is honest about its rough edges. The project is pre-1.0 (current versions are 0.9.x), and the README lists known issues:
- 2–5 second processing delay after you stop speaking before text appears. This is inherent to local model inference on CPU — not a bug, but a UX trade-off.
- First-word clipping: Occasionally, the very first word you speak gets cut off. Using push-to-talk mode (instead of toggle) reduces this.
- Whisper model crashes on certain Windows/Linux configurations. This is configuration-dependent and doesn't affect all systems. The community is actively working on fixes.
- No mobile app: Handy is desktop-only. If you need mobile dictation, use Apple Dictation (iOS) or Google Voice Typing (Android).
- No AI text cleanup: Handy outputs raw transcription. It won't remove filler words ("um," "uh"), fix formatting, or adjust tone. If you want polished output, you'll need to pair it with an LLM post-processing step — or use a commercial tool like Wispr Flow.
- Wayland support is limited on Linux, requiring
wtypeordotoolfor text input. - No file transcription or subtitle export: Handy is a live dictation tool, not a batch transcription tool.
These limitations are the price of free, open-source, fully-offline software. For most, the trade-offs are worth it.
When Should You Pay for Dictation Instead?
Free tools cover a lot of ground, but there are clear scenarios where a paid tool is the better call:
- You dictate more than 30 minutes a day and every error costs time. The 3–5% accuracy gap between free Whisper Small/Parakeet and commercial Whisper Large-v3 with custom prompts adds up to dozens of corrections per session.
- You need AI-powered text cleanup. Wispr Flow ($15/mo) and Superwhisper ($9.99/mo) both offer filler-word removal, tone adjustment, and automatic formatting. Handy gives you raw text.
- You need sub-second latency. Cloud-based tools like Wispr Flow and Deepgram process faster than local models on most hardware.
- You need mobile dictation. Handy is desktop-only. Wispr Flow, Superwhisper, and Apple Dictation all have iOS presence.
- You work in a regulated industry with specific compliance needs. Ironically, Handy's fully-offline architecture is better for privacy than any cloud tool — but if your organization requires audited, managed infrastructure, a commercial tool with SOC 2 and HIPAA compliance may be mandatory.
For everyone else — developers, writers, students, privacy-conscious users, and anyone who just wants to try dictation without a credit card — Handy is more than enough.
What This Means for You
If you're a developer or technical user, Handy is a no-brainer. It handles technical jargon far better than Google Voice Typing, it's free, you can drop in custom fine-tuned models, and the code is MIT-licensed and auditable. Install it, use Parakeet V3 for speed, and switch to Whisper Large-v3 Turbo when you need maximum accuracy.
If you're a small-business owner, Handy eliminates a recurring software cost. Dictation speeds of 125–150 words per minute (versus ~40 WPM typing) mean you can draft emails, notes, and documentation 3–4× faster — without sending a single word of confidential client communication to a third-party server.
If you're privacy-first, Handy is the only major dictation tool where you can verify the claim "your voice never leaves your device" by reading the source code yourself. No account, no telemetry, no cloud — by architecture, not by promise.
For tips on building workflows that combine AI tools productively — including voice-activated agent setups — see our guide on building a voice-activated agent OS with a free AI coder. If you're evaluating when AI tools genuinely help versus when they create risk, our piece on the deterministic-first rule for AI tools covers the decision framework.
And if you're interested in how AI agents can automate research and information gathering — the kind of work that pairs naturally with voice dictation — our automated AI research guide walks through the full pipeline.
FAQ
Q: Is Handy really completely free?
A: Yes. Handy is free and open source under the MIT license, with no paid tiers, subscriptions, word limits, or account requirements. Development is funded through GitHub Sponsors and direct donations. The entire source code is available on the GitHub repository.
Q: Does Handy work without an internet connection?
A: Yes, fully. All audio processing happens on your device using local Whisper or Parakeet models. You need internet only for the initial model download (ranging from ~478 MB for Parakeet V3 to ~1.6 GB for Whisper Large-v3 Turbo). After that, Handy works completely offline.
Q: How accurate is Handy compared to paid dictation tools?
A: On clean English audio, Whisper Large-v3 (available in Handy) achieves ~2.7% WER — comparable to or better than commercial cloud services. On real-world audio with technical jargon, Handy with Parakeet V3 correctly identified terms like "SQLite" and "LLM" in testing, while Google Docs Voice Typing garbled them. Commercial tools like Wispr Flow have a slight edge in post-processing (filler removal, formatting) but the raw transcription accuracy gap is small.
Q: What hardware do I need to run Handy?
A: For Parakeet V3 (the recommended default), you need at minimum an Intel Skylake (6th gen) processor or equivalent AMD. It runs at ~5× real-time speed on a mid-range i5, meaning transcribing 10 seconds of speech takes about 2 seconds. For Whisper models, GPU acceleration is available on Intel, AMD, and NVIDIA GPUs. Apple Silicon Macs handle all models efficiently.
Q: Can I use my own custom speech model with Handy?
A: Yes. You can drop fine-tuned Whisper GGML/GGUF model files into Handy's models directory (found in the app's data folder), restart Handy, and your custom model appears as a selectable option. This extensibility is a core design goal — the project describes itself as "the most forkable" speech-to-text app.
Q: Which platforms does Handy support?
A: Handy supports macOS (Intel and Apple Silicon), Windows (x64), and Linux (x64, Ubuntu 22.04/24.04 tested). There is no mobile app. Linux Wayland support is limited; X11 users should install xdotool and Wayland users should install wtype or dotool for text input.
Discussion
0 comments