Dictate into any app with a strict 1.35s latency deadline. Capture live dual-track meeting notes without uninvited bots. Zero focus-stealing. Pure native Swift.
Experience how Fast Talk's Cerebras inference engine instantly repairs rambling speech, removes filler words, and generates structured code, email, and meeting artifacts.
"hey let's refactor this auth controller um we should extract the bearer token from the authorization header and if it's missing return a 401 unauthorized response with a clean json body"
export function extractBearerToken(request: Request): string {
const authHeader = request.headers.get("Authorization");
if (!authHeader?.startsWith("Bearer ")) {
throw new HttpError(401, {
error: "unauthorized",
message: "Missing or invalid authorization header"
});
}
return authHeader.slice(7).trim();
}Every feature is engineered directly into macOS with private AppKit panels, Apple Neural Engine models, and SQLite WAL storage. Zero web wrappers.
Floats across all macOS Spaces and full-screen windows via NSPanel without ever stealing keyboard focus or interrupting active keystrokes.
func startDictation() {
panel.orderFrontRegardless()
audioEngine.startRecording()
}
Sub-second Cerebras gpt-oss-120b inference with strict 1.35s client cutoff. If the model or network exceeds the budget, raw text delivers instantly with zero hang.
Transforms the physical MacBook camera housing into an OS-level living status display. Physical camera exclusion geometry with a 4-tier Size Ladder (W0 Rest, W1 Glyph Wings, W2 Text Wings, W3 Expanded Card) choreographing 22 hardware states across meetings, active recording, spoken directives, and AirPods audio handoffs.
Inspects frontmost window title and terminal working directory via AXUIElement in 45ms. Switch writing styles mid-dictation or apply retroactive transformations.
Captures microphone and system audio simultaneously without sound drivers or meeting bots. Auto-detects Zoom, Google Meet, Teams, and Slack calls with EventKit heads-up.
Speak instructions naturally during meetings. "Hey Fast Talk, draft the action items" generates structured artifacts without stopping the recording.
Query meeting transcripts with complete confidence. Answers are strictly grounded in transcript evidence with verified timestamp citations.
CoreML speaker diarization runs 100% locally on Apple Neural Engine. Computes 256-d voice embeddings without sending biometrics to the cloud.
Self-updating habit ledger and custom vocabulary context terms with longest-match-first local replacement rules.
Never lose original speech. Original transcripts and audio chunks are stored in verifiable .ftsource files with SHA-256 integrity checks and 1-click restore.
"um so we need to like deploy to prod today"
Pastes cleaned text into your target app and restores your existing clipboard in 300ms. Bypasses pasteboards safely when Secure Input is active.
Instantaneous push-to-talk using modifier-only keys like Fn or Right Option. Double-tap Escape for immediate emergency cancel.
Ducks background audio to 20% or pauses MediaRemote podcasts when you dictate. Waits for audio hardware wake-up before playing cues.
Markdown-native rich text editor synchronized with audio. Option-click any word in the transcript to jump playback to that exact millisecond.
Transcribe podcasts, voice memos, and video recordings in the background directly from Finder Services or by dropping files on the Dock.
Sync notes seamlessly across your Macs via private Apple CloudKit. Share read-only snapshots with zero Fast Talk intermediate servers.
Lock sensitive notes behind Touch ID biometrics with 5-minute auto-relocking on display sleep. Automatically mirrors plaintext notes to Obsidian.
Instant access with Apple App Attest hardware verification—zero account signup required. Migrate your VoiceInk dictionary and settings in 1 click.
Most voice tools are sluggish cloud wrappers, invasive meeting bots, or unoptimized local models. Fast Talk is native macOS voice infrastructure.
| Architectural Dimension |
Fast Talk
Fast Talk for Mac
Pure Native Swift
|
Wispr Flow
$15/mo Cloud SaaS
|
Superwhisper
$249.99 Lifetime · FUP
|
Otter.ai / Fathom
Invasive Meeting Bots
|
Granola
$10/mo Manual Editor
|
Apple Dictation
macOS Built-in
|
|---|---|---|---|---|---|---|
| 1. Latency & Determinism Client AI cutoff & insertion guarantee |
1.35s Hard Cerebras Cutoff
Client cutoff with instant raw fallback. Pasting NEVER hangs.
|
3.0s – 5.5s cloud stalls
Freezes user cursor on latency spikes.
|
Variable latency
Local CPU lag on older models; cloud tier throttled under unpublished FUP.
|
5–15 min post-call
No inline cursor dictation into active apps.
|
2–5 min post-call
Post-meeting LLM summaries only.
|
~0.5s local
Verbatim-only with zero intelligent post-formatting.
|
| 2. Hardware & Space Integration Desktop presence & window activation |
Fast Notch + Non-Activating HUD
20+ Notch states + 48×8pt to 138×40pt HUD across all Spaces without stealing focus.
|
Floating pill
Steals focus; activates window away from IDE.
|
Standard window / Menu bar
No camera notch awareness; desktop window clutter.
|
Heavyweight Electron
Requires dedicated browser tab or separate window.
|
Split-screen window
Requires visible window alongside meeting; focus stealing.
|
Inline caret icon
Zero spatial HUD or hardware notch choreography.
|
| 3. Meeting Intelligence Conference audio tap & participants |
Dual-Track Native Tap (Zero Bots)
CoreAudioTapFrameSource captures mic + system audio with zero kernel drivers.
|
Microphone only
Cannot capture Zoom/Teams system audio.
|
Microphone only
No native system audio tap or meeting diarization.
|
Invasive bot participant
"Otter Notetaker has joined"; social friction.
|
Audio tap + Manual notes
Captures audio but requires user to edit notes manually.
|
None
No meeting intelligence or dual-track capture.
|
| 4. Speaker Identification Biometric voice vector handling |
100% On-Device CoreML ANE
256-d voice embeddings on Neural Engine. Zero biometrics leave Mac.
|
Cloud voice processing
Voice streams analyzed on remote servers.
|
Basic local transcribe
No speaker diarization or biometric voice matching.
|
Cloud biometric signatures
Voice fingerprints stored on vendor servers.
|
Cloud LLM separation
Audio sent to cloud for speaker turn splitting.
|
None
Single speaker only; no diarization support.
|
| 5. Context Awareness Frontmost IDE, document & terminal state |
45ms AX + 10ms Terminal CWD
Automatic style switching for Cursor, Xcode, Mail, and Terminal working dirs.
|
Clipboard context only
No terminal CWD lookup or deep AX inspection.
|
Manual mode selection
Requires manual switching between prompts.
|
Calendar title only
Zero desktop IDE or application awareness.
|
Calendar title only
No code context or terminal integration.
|
Basic caret target
Inserts raw text without IDE or repo awareness.
|
| 6. Pricing & Trust Transparency Hardware allowance & license model |
Apple App Attest Daily Free Tier
Hardware-verified allowance with ZERO signup. Transparent one-time license.
|
$15 – $23 / month
$144 – $276/yr recurring; locked 2,000 words/wk trial.
|
$249.99 lifetime / $8.49/mo
Hidden FUP throttling on cloud models; 3k trial words.
|
$10 – $30 / month
Per-user subscriptions with strict minute limits.
|
$10 / month Pro
Recurring subscription; limited free meetings.
|
Free with macOS
Lacks AI formatting, modes, and meeting tools.
|
What stays strictly on your silicon vs. what transits ephemerally under Zero Data Retention contracts.
Audio transits via encrypted WebSocket TLS directly to Soniox under strict Zero Data Retention (ZDR) and No Model Training contracts. Cloudflare AI Gateway routes the request with cf-aig-collect-log-payload: false, guaranteeing zero payload logging. Cerebras processes text cleanup under a client-enforced 1.35s timeout. Finished text injects at your active cursor in 45ms and archives to local SQLite WAL (Library.store). Zero intermediate server logs or audio retention.
All persistent audio archives, biometric voice prints, full-text indexes, and notes remain on local silicon.
Real-time processing with cryptographic Zero Data Retention and zero server logging.
Technical details on latency guarantees, bot-free audio routing, local biometrics, and zero-signup hardware attestation.
Download Fast Talk for macOS. Zero account signup. Pure Apple Silicon performance.