Said

Live captions for anything your Mac plays.

One private caption layer for calls, videos, podcasts, webinars, and voice messages.
No meeting bot. No account. No transcript history.

Download Said 0.1.0 Alpha · Build from source

_{Apple silicon · macOS 26+ · English · 4.5 MB app + one-time 731 MB model download}

Said requires macOS 26 because it captions through the system-audio tap, not the microphone.

Hear. Read. Gone.

_{Real Said captions from a Peter Yang podcast episode playing in another app.}

---

Why Said exists

The audio is already playing on your Mac, but the words are still hard to follow. Maybe the video has no captions. Maybe the call's captions are poor. Maybe the room is noisy, the volume needs to stay low, or spoken English is easier to understand when you can read it as well.

If the app does not provide good captions, the usual options are to upload a recording, invite a meeting bot, or use a separate transcript service. That is a lot of machinery for someone who just needs to read what is playing.

Said changes one thing instead: the Mac gets captions.

1. Turn Said on	2. Play anything	3. Read, then turn it off
One switch in the menu bar.	A call, video, podcast, webinar, training, or voice message.	The caption window stays where you put it. Off clears the ephemeral words and closes the surface.

No source picker. No meeting to name. No recording to clean up later.

How Said behaves

When the switch is on, Said listens only to audio macOS makes available from other processes and keeps a fixed two-line caption surface visible, even during silence. When the switch is off, capture stops and the surface disappears.

Captions are designed for reading in motion:

• A fixed left edge gives your eyes a dependable place to return to.
• Complete rows advance together instead of constantly reflowing old words.
• Stable words stay stable; only the newest tentative suffix may soften and change.
• The panel never takes keyboard focus and remains above full-screen apps.
• Hover reveals one compact bar for moving captions, changing the unified XS–XL scale, choosing Rounded/Sans/Serif/Mono/Block, and selecting white, warm yellow, or cyan text.

Why recognition stays on the Mac

Live captions require a stream of whatever you are hearing. Said keeps that stream on the Mac instead of sending calls, videos, and voice messages to a transcription server.

After one verified model download, speech recognition runs locally with Parakeet Unified EN 0.6B Q8_0 on Metal.

Said can observe	Said does not collect
Audio playing through the supported Mac system-audio path while Said is on	Microphone input
Temporary PCM and caption text in process memory	Screen pixels, camera, keyboard, clipboard, files, browser history, or window titles
Caption appearance and placement preferences	Audio recordings or transcript history
Model/update network requests	Accounts, analytics, cloud inference, or remote caption logs

Said persists the verified speech model, its content-free installation receipt, and ordinary preferences. It does not claim cryptographic memory erasure; audio and caption buffers are released when the stream ends or the app quits.

Read the complete privacy contract, threat model, and verification matrix.

Install the alpha

Said currently requires an Apple-silicon Mac running macOS 26 or later.

1. Download Said-0.1.0-alpha.dmg.
2. Open the DMG and drag Said into Applications.
3. Control-click Said, choose Open, then confirm.
4. Allow System Audio Recording Only when macOS asks.
5. Let Said download and verify its English speech model once.
6. Play something with spoken English and turn the Said switch on.

Important

This is a playable, ad-hoc-signed alpha. It is not yet a notarized public release. The Control-click → Open step is required because no Developer ID certificate is configured yet. Do not redistribute this alpha as a finished release.

The release page publishes the DMG checksum and build commit: Said 0.1.0 Alpha.

Scope

Said is	Said is not
A native caption layer for Mac system audio	A meeting recorder or transcript archive
English-only in V1	Translation or multilingual recognition
One opinionated local model	A model picker or cloud fallback router
A deliberate on/off mode	An ambient surveillance process
A floating reading surface	A dashboard, chat box, or AI workspace

There is intentionally no microphone capture, rewind, save, export, search, speaker labeling, summary, action item, calendar integration, subscription, analytics SDK, or account system.

Every proposed feature has to improve one loop: play audio, read captions.

Implementation

Said is a Swift 6 application with no browser shell, Python service, local HTTP server, model daemon, or cloud recognizer.

Mac process audio
    ↓
Core Audio process tap (Said excludes its own audio)
    ↓
48 kHz stereo → 16 kHz mono Float32, bounded in memory
    ↓
Parakeet Unified EN 0.6B Q8_0 through pinned transcribe.cpp
    ↓
Metal-backed stable-prefix streaming
    ↓
AppKit NSPanel + SwiftUI caption content

The code keeps capture, normalization, recognition, model management, caption state, and UI independently testable:

SaidApp       AppKit lifecycle, menu bar, setup, settings, caption panel
SaidCapture   Core Audio capture, normalization, bounded PCM pipeline
SaidASR       Parakeet streaming actor and transcribe.cpp adapter
SaidModel     Resumable download, SHA-256 verification, atomic install
SaidCore      Pure caption, state, diagnostics, and buffer behavior

Dependencies and model artifacts are pinned by immutable revision, exact byte count, and SHA-256. Start with the architecture, model provenance, and product requirements.

Verification

The repository separates deterministic CI from tests that require a physical Mac, real system audio, or the 731 MB model.

Verified today:

• Real Mac system audio reaching the local Parakeet caption pipeline.
• Metal model loading and true streaming.
• 45 deterministic tests plus the static privacy gate.
• Regression coverage for caption stability, bounded PCM queues, capture recovery, model verification, and resumable downloads.
• GitHub Actions building and verifying an arm64 app without bundling the model.
• Content-free performance, privacy, and release receipts under Artifacts/Receipts.

The release checklist clearly separates what is proven from what still blocks a finished public release: physical M1/16 GB performance, long-session receipts, final model-license review, Developer ID signing, notarization, and a clean-machine install.

Verification commands

# Deterministic tests and static privacy contract
./scripts/run-tests.sh

# Real model streaming
./scripts/streaming-smoke.sh

# Real Mac playback → local model → captions
./scripts/capture-smoke.sh

# Installed model + per-process network observation
./scripts/offline-smoke.sh

# Known marker + quit + retained-content inspection
./scripts/post-quit-privacy-smoke.sh

# Long-session stability and bounded memory
./scripts/soak-smoke.sh --minutes 30

# Performance and privacy receipts
./scripts/performance-smoke.sh
./scripts/privacy-smoke.sh

Build from source

You need Xcode with the macOS 26 SDK on an Apple-silicon Mac.

git clone --recurse-submodules https://github.com/r3dbars/said.git
cd said
./scripts/bootstrap.sh
./scripts/build-and-run.sh --verify

After the first build, reopen the same bundle without replacing its local code-signing identity:

./scripts/build-and-run.sh --launch-only

Local development builds are ad-hoc signed unless you provide a trusted signing identity. Rebuilding an ad-hoc bundle can make macOS ask for System Audio permission again.

Contributing

Focused issues and pull requests are welcome. Please read CONTRIBUTING.md before changing capture, inference, privacy, or product scope.

Before opening a pull request:

./scripts/run-tests.sh
./scripts/privacy-smoke.sh

Hardware-only behavior needs a hardware receipt. A unit test is not proof of a macOS permission, Core Audio capture, Metal latency, power, or thermal claim.

Repository guide

Path	Purpose
Sources/	Native application and independently testable modules
Tests/	Deterministic caption, capture, model, and privacy-adjacent tests
scripts/	Bootstrap, build, smoke, soak, privacy, and release tooling
docs/PRD.md	Locked product contract and acceptance bar
docs/model-spike.md	Streaming benchmark and runtime evidence
docs/reliability-matrix.md	Proven, partial, and open verification gates
docs/release-checklist.md	Alpha and public-release requirements

License

Said application code is available under the MIT License. Runtime and model artifacts retain their own licenses and notices; see third-party licenses and model provenance.

Public distribution remains blocked until the pinned model/conversion license chain receives explicit human review.