Plainsay

Hold a key. Speak. Release. Get text at your cursor—or safely on the clipboard.

A native Mac dictation app in a 12 MB download, with Whisper and Parakeet running locally.

Star Plainsay on GitHub ★ · Download Plainsay.dmg · Homebrew · Website · Benchmark

12 MB native app · free Local mode · no account · no telemetry

Plainsay is free and MIT-licensed. With Local transcription and Polishing off or running on your Mac, dictation audio is not uploaded, no account or subscription is needed, and recognition works offline after the model download. Automatic checks for signed updates are enabled by default and can be disabled in Settings. Optional Plainsay Cloud provides hosted transcription and Polishing for US$4 gross/month (VAT included); Cloud speech sends recorded audio to the service and includes up to 900 transcription minutes in any rolling 30-day window. An active subscription also unlocks Auto-translate and automatic email layout for new dictations from any transcription source. In v0.2.29, those styles require a BYOK or compatible local Polishing provider; the built-in Plainsay Cloud Polishing request does not yet carry their instructions.

Requirements: macOS 14 or newer and an Apple-silicon Mac. The app is signed and notarized. Recommended multilingual models are separate one-time downloads of approximately 475–632 MB; smaller English-only options start at about 150 MB.

Install

Direct download

Download the latest notarized DMG and drag Plainsay to Applications. The stable link always points to the latest release; the release page has checksums and release notes.

Homebrew

brew install --cask conrader/plainsay/plainsay

Plainsay checks for signed updates automatically. You can also use Check for Updates… from its menu-bar menu.

If Plainsay earns a place in your menu bar, star the repository. It is the simplest way to help more Mac users find the project.

Why Plainsay?

• Small, native Mac app. The current notarized DMG is about 12 MB. Plainsay is written in Swift and stays out of the way in the menu bar.
• Two genuinely local engines. Choose Whisper through WhisperKit or multilingual NVIDIA Parakeet through FluidAudio.
• Measured, not hand-waved. The committed benchmark records every reference, hypothesis, and timing. On the benchmark machine and its disclosed 50-clip English sample, prepared-model average latency was 0.10 s for Parakeet and 0.68 s for Whisper, with the first inference included.
• Broad insertion support. Plainsay uses the standard paste flow across many native apps, Electron apps, web views, and terminals, then attempts to restore the previous clipboard contents.
• Safe fallback. Polishing is optional and time-bounded. If it is disabled, offline, or fails, Plainsay continues with the raw transcript. If no safe paste target is found, the text stays on the clipboard.
• Clear open-source boundary. The Mac client code, benchmark harness, and release scripts are MIT-licensed and auditable; the committed LibriSpeech clips are CC BY 4.0. Plainsay Cloud is a separate optional hosted service.

Verifying the models

Plainsay is signed and notarised, then downloads a Core ML speech model at first run and executes it. Neither WhisperKit nor FluidAudio verifies a file it has just fetched, so until v0.2.27 nothing did. Expected digests for every supported model now ship inside the signed bundle. Before loading, Plainsay hashes the pinned files that were actually downloaded and refuses any present file whose contents differ. Repository paths absent because a library downloads only a subset are reported, not treated as tampering.

Nobody was checking the speech models — what the libraries actually guarantee, why verifying against a hash from the same server proves nothing, and what the fix does and does not buy.

How it compares

Side-by-side pages for the alternatives, written to be useful rather than flattering — each one ends with the cases where the other app is the better buy.

All the main Mac dictation apps compared — price, what stays on your Mac, what needs an account.

vs Wispr Flow · vs superwhisper · vs VoiceInk · vs MacWhisper · vs Apple Dictation

One worth stating here: Plainsay dictates, and that is all it does. It has no file import, so if you have recordings to transcribe rather than words to speak, MacWhisper is the tool for that job.

What leaves your Mac?

The hosted routes are explicit:

• Your own provider: audio, a language hint, and optional vocabulary prompt go directly to the speech endpoint you configure (Groq, deAPI, OpenAI, or a compatible custom service). Hosted Polishing sends transcript text, vocabulary terms, and any enabled translation-target or email-layout instructions directly to Google Gemini, Anthropic, OpenRouter, OpenAI, or the compatible endpoint you configure.
• Plainsay Cloud: encoded audio, clip duration, language, and an optional vocabulary prompt go to api.plainsay.app. Speech may be processed on a Plainsay-operated transcription node or forwarded to deAPI. If Cloud Polishing is enabled, transcript text and vocabulary terms are sent from Plainsay Cloud to OpenRouter, using its configured Gemini model. The v0.2.29 Cloud cleanup request does not send the Auto-translate target or email-layout instruction.
• Optional Voice Filter: filtering runs on the Mac, but it is best-effort. If the filter is unavailable or fails, a hosted speech mode receives the unfiltered recording.

Plainsay Cloud's application database stores account, authentication, subscription, and usage records: an email address and/or Apple account identifier, session records, Stripe customer/subscription identifiers and status, plus each operation's timestamp, type, audio duration, estimated cost, and Polishing token counts. It deliberately stores neither recorded audio nor transcript text. Stripe receives the account email and handles checkout and payment details; Apple handles Apple sign-in, and the configured email-delivery service receives the address and sign-in message.

API proxy access logging is disabled. Minimal application event logs contain only the method, matched route template, status, and elapsed time; they exclude network addresses, headers, query strings, and request bodies. Website and signed-update delivery can still produce ordinary web-server access records.

Local mode does not mean the app never uses the network: the selected model must be downloaded once, and the default update check contacts api.plainsay.app (it can be switched off). The Mac client sends no analytics or telemetry. By default, History keeps up to 100 raw and final transcript records on this Mac for 30 days. You can choose 7, 30, 90, or 365 days, clear the records, or stop saving them; turning History off deletes saved transcripts. Audio is staged in an owner-only local folder while a dictation is processed and removed after a successful or completed attempt. Interrupted recordings can remain for recovery. On each launch, Plainsay removes ones older than 7 days and keeps no more than the newest 20; clearing History purges them too. These files are excluded from backups.

API keys and the Plainsay Cloud session token are stored in macOS Keychain. Local mode is the default choice on a new installation; cloud use requires an explicit choice.

First dictation

1. Choose On this Mac for Local mode, then pick the recommended model for your languages. Choose a main language, then add any others you use. The main language is sent to providers and models that accept only one language hint, and is Whisper's fallback when a retry is needed.
2. Grant Microphone, Accessibility, and Input Monitoring permissions. Input Monitoring lets Plainsay notice the configured dictation shortcut, the ⌃⌥⌘T translation toggle, and Escape while recording; it never records what you type or includes keystrokes in a transcript. On the first request, Plainsay brings the system prompt to the front. If you request the permission again after macOS no longer has a prompt to show, Plainsay opens the correct System Settings pane. macOS may quit Plainsay after a new grant takes effect; reopen it and Setup resumes at the same step.
3. Hold Right Command, speak, and release. Tap it once instead if you prefer toggle mode; tap again to stop.

Each dictation can run for up to 10 minutes. At the boundary, Plainsay stops recording automatically, shows a notice, and processes the captured audio.

The shortcut, languages, vocabulary, model, and hold/toggle behavior can all be changed later in Settings.

Languages, translation, and email layout

• Wrong-language guard. With spoken languages configured, Whisper retries an evidently out-of-list result in your main language. Parakeet cannot force that retry, so it records narrow, high-confidence mismatches in diagnostics instead of claiming to have fixed them.
• Auto-translate (Cloud entitlement). Choose a target from your configured spoken languages or English in the menu-bar menu. With a style-aware BYOK or compatible local Polishing provider selected, each new dictation is translated until you switch it off. Press Control–Option–Command–T (⌃⌥⌘T) anywhere to toggle the last target. Translation is off by default and requires Polishing to be enabled and configured.
• Email mode (Cloud entitlement). This is off by default. Enable Format dictation as an email in Settings › General. In supported mail apps and recognized webmail compose windows, Plainsay separates a spoken greeting, body paragraphs, and sign-off without adding words you did not say. It also requires Polishing.
• Current provider limitation. In v0.2.29, the style prompt is honored by BYOK providers and compatible local endpoints. The built-in Plainsay Cloud Polishing route still performs ordinary cleanup because its client request omits the translation and email-layout instructions. Audio recovered after a relaunch is also polished as plain text.
• Vocabulary suggestions. Settings can propose recurring names and product terms from saved dictations. Nothing is added unless you choose Add, and you can dismiss a suggestion instead. Applied vocabulary corrections use conservative edit thresholds, score every candidate, and choose the closest match rather than the first acceptable one.

How it works

key down ──► remember frontmost app → show HUD → record 16 kHz mono audio
key up ────► transcribe locally, in Plainsay Cloud, or through your API
             optional Polishing, translation, and email layout
             (bounded timeout or truncated reply → raw-transcript fallback)
             snapshot clipboard → paste text → attempt clipboard restore

The local engines run as Core ML models accelerated on Apple silicon. Polishing asks the selected model to turn spoken language into written language while preserving wording, meaning, and tone; by default it also preserves the language. With Auto-translate enabled and a style-aware BYOK or compatible local provider selected, it renders each new result in the target language instead. Review important text. Polishing can use Plainsay Cloud, your provider, a compatible local endpoint, or remain off; the current Cloud request performs ordinary cleanup without the translation or email-layout style. It is also instructed not to invent an ending for a transcript cut off mid-sentence; if a provider reports that its own reply hit an output limit, Plainsay keeps the raw transcript instead of inserting partial edited text.

Read How it works for the pipeline and Why on-device for the privacy trade-offs.

Reproducible benchmark

These are prepared-model results on a disclosed 50-utterance sample from LibriSpeech test-clean; the first inference is included and no hidden transcription warm-up is subtracted. They are not a claim about every language or noisy real-world dictation. The exact Mac, source commit, dependency pins, methodology, limitations, raw results, and one-command harness are in BENCHMARK.md.

Screenshots

_{The listening HUD stays beside the text cursor while a dictation is active.}

Troubleshooting

Settings › About shows the installed version and has a one-click Copy Diagnostics Command action. Run the copied command in Terminal when a dictation, permission, model, or provider problem needs a detailed report. Its audio diagnostics compare how long recording was active with how much audio was actually captured, making an early microphone cut-off visible.

Build and contribute

Bug reports, small fixes, tests, documentation improvements, and focused feature proposals are welcome. Start with CONTRIBUTING.md, which covers the development build, tests, and how to propose a change.

swift build
swift test

For security issues, please follow SECURITY.md instead of opening a public issue.

Contact

• Bugs: open an issue
• Ideas and questions: GitHub Discussions
• Everything else: hi@plainsay.app