Talk through what's on your Mac.
Push-to-talk from any app. On αlpha reads terminal output, selected text, cursor context, PR diffs, browser state, and Slack threads, then translates when needed and talks back with tool-augmented answers in voice.
Use Wispr Flow to write. Use On αlpha to understand and act on what you're looking at.
What is On αlpha?
On αlpha (音α) is a spoken reasoning layer for macOS, built by ORBWEVA Labs. It is not a dictation app or a translation app. Press a hotkey from your editor, terminal, browser, Slack, or GitHub window and ask about what is already on screen: a stack trace, selected code, a PR diff, an English doc, or a Japanese comment. On αlpha reads that context, translates only what matters, reasons with tools like web research, Claude Code handoff, and clipboard operations, then answers back in voice in under a second. The first wedge is builders and bilingual operators who already work across English-first software and local-language thinking. English, Japanese, and Korean are supported from day one. Currently in a 90-day public build from May to August 2026, with founding-member pricing reserved for the first 100 waitlist signups.
Built by ORBWEVA Labs, led by Ryan Ahamer, founder of two Japanese kabushiki kaisha: LokaLingo KK and Accent Language KK.
Your voice doesn't leave your Mac.
Defensible Mode and on-device ASR ship in v1.1 (weeks 5–7 of the 90-day arc). The cloud-Realtime prototype runs the loop in weeks 1–4; audio cuts over to local Whisper.cpp by week 7.
Read the privacy postureOn αlpha runs speech recognition on-device with Whisper.cpp. Audio never reaches a server for transcription. The reasoning step (cloud LLM) receives text and screen context, not audio. The synthesis step (ElevenLabs) receives text and returns audio. Your voice is one-way: in, then forgotten by the system the moment a token is emitted.
Architectural, not contractual.
Wispr Flow promises not to train on your data. On αlpha cannot send it in the first place. Verify with Little Snitch, Wireshark, or your network tab.
Your voice profile lives on your Mac.
On αlpha learns your accent, your terminology, your code-switching. The adapter weights sit in ~/Library/Application Support. Never synced. One click to forget.
Defensible Mode is the default.
Cloud reasoning and telemetry are explicit opt-in per call. Turn the whole loop local for short queries. Turn it off entirely to audit egress.
The video will show one loop.
The recorded demo will replace this panel. For now, the page shows the exact interaction visitors should understand: hotkey, screen context, tool reasoning, spoken answer, next action.
Hotkey
Reads screen
Uses tools
Talks back
Next action
"Why is this terminal error happening, and what should I check next?"
It reads the stack trace, notices the selected file, checks the likely package issue, and talks through the next file and command to try.
The real video should use a terminal or PR-review scene, not a generic assistant chat.
How does On αlpha work?
Hears you
Push-to-talk from any app. Reads your cursor, last terminal output, selected text. Knows what you're looking at.
Reasons with tools
GPT-Realtime-2 native speech-to-speech. Perplexity research, web-fetch, Claude Code handoff, clipboard ops.
Talks back fast
Sub-second first audio byte. Cedar and Marin voices included, or clone your own via ElevenLabs PVC. Natural barge-in.
Built for moments already on screen.
On αlpha should prove itself in specific work moments, not as a vague assistant.
Debugging
Ask about the terminal error in front of you. On αlpha reads the stack trace and selected file, then talks through the likely cause and next command.
PR review
Ask for a spoken risk check before merge. On αlpha reads the selected diff and explains what changed, what looks risky, and what needs another look.
Technical interpretation
Ask in Japanese, Korean, or English about docs, comments, tickets, or errors. On αlpha translates only what matters, then explains the technical context in the language you are thinking in.
A layer on top of your stack.
On αlpha does not ask you to abandon dictation, coding agents, or chat. It gives those workflows a spoken screen-aware layer.
Complement, not compete.
Wispr Flow and Willow handle voice in. Apple handles OS-level assistant tasks. On αlpha is the spoken reasoning layer for active work on screen.
| On αlpha | Wispr Flow | Apple Intelligence | |
|---|---|---|---|
| Best for | Talking through active work | Dictating into fields | General assistant tasks |
| On-device ASR | Default. Audio stays on your Mac. | No. Cloud transcription, zero-retention contract. | Yes. But macOS-locked and no tools. |
| Direction | Voice out + tools | Voice in | Assistant / OS actions |
| Languages | EN / JA / KO from day 1 | English only | Limited JA / KO |
| Tools | Web, code, clipboard, handoff | None | Apple ecosystem |
| Translation | Technical interpretation | Text cleanup | System translation |
| Latency | < 1s first audio byte | n/a (text out) | Variable |
| Voice cloning | ElevenLabs PVC (Studio) | No | No |
| Reads your screen | Cursor, terminal, selection | No | App-scoped |
Your voice for macOS.
- Reads terminal output, selected code, browser context, and Slack threads
- Answers in voice with web, code, and clipboard tools
- Translates when needed, then explains what matters in Japanese, Korean, or English
- Keeps the keyboard free for code, not prompt typing
Builders working across context.
- Wispr Flow users who want voice out, not another dictation app
- Developers reading a PR diff or terminal error with headphones in
- Bilingual operators crossing Japanese, Korean, and English software work
- Anyone who said "I wish I could just ask my Mac"
Founding-cohort pricing.
Founding members get 50% off the BYOK Pro and Studio tiers for 12 months, plus 30-day money back if the loop doesn't ship under one second on your Mac. macOS first; Windows coming later. Founding pricing covers both platforms.
- Free: five voice turns a day, all tools, no memory.
- Pro and Studio: BYOK on your OpenAI key, unlimited usage, voice cloning on Studio.
- Team: per-seat with SSO and audit.
When does On αlpha launch?
Build started May 2026. Day-14 GO/KILL falls on Week 2. macOS first; Windows committed for 2027.
Frequently asked questions
How is On αlpha different from dictation tools like Wispr Flow?
Use both. Wispr Flow is voice in: you speak, it types. On αlpha is voice out: you speak about what is already on screen, it reasons with tools, and it answers back in voice in under a second. Dictation writes into fields. On αlpha talks through active work.
Does On αlpha work in Japanese and Korean?
Yes. English, Japanese, and Korean are supported from day one, not as a localization phase. The model speaks back in the language you spoke in, including mid-sentence language switches. Built by a Japan-resident team and tuned for bilingual knowledge workers.
Is my voice data private? Where is it processed?
Audio doesn't leave your Mac for transcription. Whisper.cpp runs locally on Apple Silicon (v1.1+, shipping in weeks 5–7 of the 90-day build). The text transcript reaches OpenAI Realtime for the reasoning step and ElevenLabs for the synthesis step — both gated by Defensible Mode and opt-in per call. No raw audio is stored by ORBWEVA Labs, ever. See the privacy posture page for the full data flow.
Is there a Windows version of On αlpha?
macOS first. Windows is committed for 2027 and founding-member pricing covers both platforms. The macOS desktop scaffold ships in Weeks 1–4 of the build; the Windows port follows after public launch.
What is the Day-14 GO/KILL gate?
Two weeks after the build starts, On αlpha needs at least 100 waitlist signups OR 10 "would pay" replies to trigger the full 90-day commit. If neither threshold lands, ORBWEVA Labs publishes a public kill report explaining what didn't work and reallocates effort. No quiet pivots, no sunk-cost development.
Can I clone my own voice in On αlpha?
Yes, on the Studio tier. Voice cloning uses ElevenLabs Professional Voice Cloning (PVC). Record a short sample, get a model trained on your voice, and have On αlpha respond in it. Cedar and Marin (OpenAI's stock voices) are included on all paid tiers as the default.
Does On αlpha work offline?
Speech recognition does, in v1.1+. Whisper.cpp runs locally on your Mac, so push-to-talk and transcription work without a network. The reasoning step (OpenAI Realtime) and synthesis (ElevenLabs) need a connection. Defensible Mode lets you cap a query at the local layer when offline.
When can I try On αlpha?
Public launch is around 90 days after the May 2026 build start, with Product Hunt timed for Week 13. The first 10 invited beta users land in Week 8 against a live Stripe integration. Founding-member pricing is reserved for the first 100 waitlist signups.
Day 14: we ship or we publish the kill report.
Be the first to try On αlpha.
Drop your email and one line on what you'd use it for. Founding-member pricing reserved for the first 100 signups.