OLIV: Thai speech-to-text and voice typing for Mac

Hold a key, speak Thai mixed with English, and OLIV types it into any app.
Offline dictation for Apple Silicon. Free and open source.

ภาษาไทย · Benchmark · OLIV vs Wispr Flow · Changelog

Thai speakers who work in tech mix languages constantly: "ตัว build fail เพราะ dependency version ไม่ตรง". Most Thai speech-to-text spells those English words out in Thai script, so you get:

ตัวบิวท์เฟลล์เพราะดิเพนเดนซี่เวอร์ชั่นไม่ตรง

Then you fix every word by hand. OLIV turns the English words back into English, leaves the Thai alone, and pastes the result at your cursor. Speech recognition and cleanup run locally on your Mac by default.

• 🇹🇭 Built for Thai + English. Uses Typhoon Whisper Turbo, a Thai fine-tune of Whisper, plus a cleanup pass that restores English tech words.
• 🔒 Local by default. After a one-time model download, dictation works offline. Your audio stays on your Mac unless you turn on a cloud engine yourself.
• ⚡ About a second. Median 1.0 s from releasing the key to text, measured on an M4 Pro.
• ⌨️ Works in any app. Slack, VS Code, Terminal with Claude Code, Notes, a browser: OLIV pastes wherever your cursor is.
• 📖 Learns your words. Add names and jargon to your Dictionary and OLIV spells them your way, even when the speech model gets a Thai name slightly wrong.
• 🔢 Formats Thai. สี่สิบห้า becomes 45, สองจุดสี่จุดหนึ่ง becomes 2.4.1, and มากมาก becomes มากๆ.
• 🎧 Ignores your music. Echo cancellation subtracts what your Mac is playing, so songs don't turn into text.
• 🆓 Free, MIT-licensed, no account. No subscription and no telemetry.

1. Speak. Hold the push-to-talk key (Right ⌥ by default). A small waveform pill shows that OLIV is listening, and your words appear in it as you speak.
2. Transcribe. Typhoon Whisper Turbo runs on the GPU with MLX.
3. Fix known words. A Rust core applies a built-in dictionary, sound-alike matching, number and ๆ formatting, and your personal Dictionary.
4. Restore English. A small on-device language model (Gemma E2B) changes Thai-spelled jargon back to English and leaves everything else as you said it.
5. Paste. The text appears at your cursor with ⌘V. Per-app verbatim mode skips cleanup for terminals and password managers.

• The cleanup step makes the difference. On the held-out set, accuracy drops to 60% without it.
• OLIV's raw speech model is not the strongest. Whisper large-v3 transcribes slightly better before cleanup. OLIV wins on the full pipeline while staying local and about half the size. All results were re-measured on 0.2.2 (28 October 2026).
• Names are tested on a public benchmark too. On Thai ContextASR Bench (2,062 utterances from many voices), OLIV 0.2.2 recognizes 63% of the names and terms on a supplied list, up from 46% before. See the method and results.
• These numbers have limits. "Meaning match" is LaBSE semantic similarity of at least 0.80, not word-for-word accuracy. All clips were recorded by one speaker (the developer). On FLEURS Thai, read by many speakers, 95% of recordings kept their meaning, though the exact wording varied with the voice (details). Recordings by other people on their own Macs and in noisy rooms are still to come. Names and numbers can be wrong even when the meaning counts as a match, so check before you send.

The interactive benchmark includes failures, playable audio and the full method.

0.2.2: your Dictionary understands Thai spellings. Thai names and terms in your Dictionary now fix near-miss spellings (บางแคร์ → บางแค, พงศ์ → พงษ์), and a name you add in English replaces its Thai spelling (คลองเตย → Khlong Toei). Long lists cause fewer wrong substitutions, and dictation with an empty Dictionary is unchanged.

0.2.1: one Dictionary, synced. Vocabulary and Replacements are now a single Dictionary with stars, on/off switches, search and CSV import/export. With the OLIV API it syncs across your devices, and OLIV can suggest words you keep correcting. Nothing is added until you approve it.

0.2.0: native Rust and MLX.

OLIV 0.2.0 removed Python from the app. Text processing and API requests use a shared Rust core, and Whisper and Gemma run on native MLX in Swift. Accuracy stayed on par with 0.1.x (mean WER 12.34% vs 12.58% on the main corpus). It also brought a redesigned Settings window, faster push-to-talk start-up and the opt-in OLIV API engine. See the changelog for every release and the migration notes.

1. Download OLIV.dmg (27 MB, signed and notarized by Apple).

2. Open it and drag OLIV into Applications.

3. Launch OLIV and allow Microphone and Accessibility access. Accessibility lets OLIV paste into other apps.

4. On first run, OLIV downloads its models from Hugging Face, about 5 GB once:

   • Speech: chayapats/typhoon-whisper-turbo-mlx, about 1.5 GB (an MLX conversion of SCB 10X's Typhoon)
   • Cleanup: mlx-community/gemma-4-e2b-it-4bit, about 3.3 GB

   Don't want to wait? In Setup, Try it now through the OLIV server starts a free trial: until the download finishes, your recordings are sent to the OLIV server for transcription (your Dictionary isn't). When the models are ready, OLIV switches back to your Mac and tells you. It's off unless you choose it, and only builds that come with a server (like the DMG above) offer it.

5. Hold Right ⌥ Option, speak, and let go.

Requirements: an Apple Silicon Mac (M1 or newer) with macOS 14 Sonoma or later. The OLIV API engine needs an internet connection and an API key instead of local models.

All settings live in the menu bar: click the OLIV icon, then Settings….

1. Open Dictation engine.
2. Keep Typhoon Whisper Turbo selected. It's the default, so a fresh install needs no changes.
3. Check that the banner says Audio stays on this Mac. Leave Text cleanup on for the full 98%.

After the one-time model download (about 5 GB), dictation works without internet and your audio is never uploaded.

For developers and teams. Run speech-to-text and cleanup on a Linux machine with an NVIDIA GPU, then point your Mac, Linux and Windows computers at it. Set up the server with server/README.md and create a key for each device with oliv-keys. Then, on the Mac:

1. In Cloud services, turn on Enable OLIV API.
2. Enter your Server URL, the base URL only (for example https://oliv.example.com).
3. Paste this device's Individual API key. It's stored in the macOS Keychain.
4. Click Check connection.
5. In Dictation engine, choose OLIV API.

The Mac downloads no models in this mode. Audio goes only to your server and to whatever terminates TLS in front of it, and your Dictionary syncs across the devices whose keys you link with oliv-keys profile link (each key has its own Dictionary until linked; see server/docs/DICTIONARY.md). On Linux and Windows, use the OLIV desktop client.

Groq runs Whisper large-v3 in the cloud. OLIV can still clean up the text on your Mac.

1. Create an API key at console.groq.com. In Cloud services, turn on Enable Groq.
2. Paste the key. It's stored in the macOS Keychain.
3. In Dictation engine, choose Groq.
4. In Text cleanup, choose how much OLIV does after Groq:
   • On (91% meaning match): Gemma runs on your Mac (about 3.3 GB download), restores English words, formats Thai numbers and applies your whole Dictionary.
   • Off (85%): no model to download. Only Dictionary entries with Correct a misspelling are applied. Plain Dictionary words are sent to Groq as recognition hints.

Your audio goes to Groq while this engine is selected. If Groq fails, OLIV retries once on Typhoon Whisper Turbo, but only if that model is downloaded.

• With the default on-device engines, audio is processed on your Mac and is never uploaded.
• Cloud engines are off until you enable one, add a key and select it. The one shortcut is the free trial you can choose while the models first download; it's labelled Cloud and ends by itself when they're ready.
• Recent dictations are kept in memory only and are cleared when you quit.
• Names on screen (off by default) reads the focused window's text through Accessibility when you start dictating. The text is used for that one dictation, kept in memory only, never saved or logged, and never leaves your Mac, even with a cloud engine.
• Copy Diagnostics creates a plain-text support report that never includes transcripts or API keys.
• Report a Wrong Result (Mac: Recent… ▸ Report a Wrong Result; iPhone: History) opens a preview you can edit, then a prefilled public GitHub issue in your browser. The report holds what OLIV typed, the raw transcript, what you meant (if you add it), the engine and whether cleanup ran, the app and OS version, and timing. It never includes audio, your server address or API keys. OLIV sends nothing itself: the issue is posted only if you submit it on GitHub.
• OLIV has no analytics and no telemetry. The network is used for the model download, update checks and any cloud engine you choose.

Requires an Apple Silicon Mac with macOS 14+, Xcode, Rust and XcodeGen. Python is not needed.

brew install xcodegen ripgrep
xcodebuild -downloadComponent MetalToolchain
bash scripts/build_app.sh   # → build/OLIV.app
bash scripts/test.sh        # Rust, native runtime and Swift tests

Repository layout:

See CONTRIBUTING.md and benchmark/README.md to run the benchmark on your own recordings.

Issues and pull requests are welcome, in Thai or English. Please read CONTRIBUTING.md and the Code of Conduct. Report security issues privately as described in SECURITY.md.

If OLIV saves you time, a ⭐ helps other Thai speakers find it.

• Code: MIT © 2026 Chayapat Sriwattanachote.
• Speech model: Typhoon Whisper Turbo by SCB 10X (MIT), based on OpenAI Whisper. OLIV ships an MLX conversion; credit for the model goes to the original authors.
• Cleanup model: Gemma, under its own upstream license.
• Libraries: MLX Swift, Sparkle and the Rust crates listed in rust/Cargo.lock. Their licenses ship inside the app.

Third-party names such as Groq, Whisper and Wispr Flow belong to their owners. Comparisons use our own Thai–English dictation set and were run once.