Skip to content
tensorhaze

Menu

Motion

Reduced motion shows still images instead of animation. Everything else stays the same.

Tool 1 · Speech to text

Amethyst

Turns speech into clean, readable text.

  • Invitation
  • Mac, Web live · iOS beta
Open AmethystSign in with an invite key, then a passkey.
Sample

One take, four views

um so i think we should uh move the review to thursday because the the build isnt green yet and the owner is out

Raw: what the recognition step returned, with fillers, repeats and no punctuation.

Any view can be translated into another language.

01

What it does

You speak or record. Amethyst transcribes it, and an optional editing pass proofreads, polishes or rewrites the take.

  • Dictate into any fieldOn a Mac, press ⌃⌥D, speak, and press it again. The text lands in the field you were typing in and on the clipboard. A helper app can also deliver it to a field on a remote Mac.
  • Four views of a takeEvery take keeps its Raw text. From it you get Proofread (light, standard or aggressive), Polished and Prompt, and any view can be translated.
  • Record long sessionsIn the web app, a take over 8 MB (about 33 minutes) uploads in resumable 4 MB parts, so a dropped connection costs one part, not the take. One recording can be up to 30 MB, about two hours.
  • Live captureWatch words appear while you speak. There are three latency presets: Fast (about 1.2 seconds), Balanced (about 1.5) and Accurate (about 2.5). These are preset settings, not measured delays.

02

Where it runs

Six ways in. They share one library of takes.

EditionRuns onState
Mac menu-bar appApple Silicon, macOS 14 or laterVersion 1.20.0, live. Updates through Sparkle.
iPhone and iPadOne universal appBeta, build 22 (18 Sep 2026)
Apple WatchComes with the iPhone appBeta, in build 22
Web appAny browser, installableLive
iPhone ShortcutsAmethyst Quick Copy and Amethyst DictateInstalled from a public Shortcuts page
HTTP APIPOST /api/v1/transcribe with a Bearer tokenToken required

03

Vocabulary

Teach it the names and terms it should never get wrong. Three lists, plus word packs and a built-in IT glossary.

Sample entries in the format the app expects

Personal terms · one per line

Kubernetes
WebAuthn

Anti-correct · never modified by proofread

codename-onyx

Replacements · "heard = written", one per line

post gress = Postgres

04

One library

Every take from every client lands in one history.

  • SearchFull-text search across your whole history, with a filter by the device a take came from.
  • Eight export formatsMessage, List, Note, Email, Markdown, JSON, Plain text and PDF.
  • Share links that expireA link lasts 7, 30 or 365 days, or has no expiry, and its creator can revoke it. The shared page asks search engines not to index it.

05

Where your data lives

Audio and text are sent to Groq for recognition and editing. Text stays on the Amethyst server until you delete it, and audio is deleted after 90 days by default.

Third party
Audio goes to Groq for recognition, and the text goes to Groq for editing. Amethyst needs an internet connection.
Recognition
A fast Whisper model by default. A slower, more accurate Whisper model when you re-run a take.
Stored audio
Deleted after 90 days by default.
Text and history
Kept on the server in a SQLite database with full-text search, until you delete a take.
In your browser
A long recording is parked in the browser until its upload completes, so a reload can resume it.
Share links
The server stores only a hash of the link's token.
Access
By invitation. Sign-in is an invite key, then a passkey.

06

Help and access

Who can use it. Amethyst is by invitation. You sign in with the invite key the lab gave you, then save a passkey on your device; after that the passkey is all you need.

Where Open goes. Open takes you to Amethyst's sign-in page. Once you are signed in you land in the web app. The Mac, iPhone and Watch apps are installed separately and use the same account.

If it does not open. The web app runs on the lab's own server. If the page does not load, try again later. When the lab knows the service is down, this page says so at the top.

07

Sources

Built with

Whisper on Groq, a FastAPI backend and SwiftUI apps.

  1. Lab project plan, 2026: the status of each platform.
  2. Lab web service notes and configuration, 2026: speech recognition on Groq, the FastAPI backend and the sign-in.

Facts as of 2 Oct 2026.