OLIV

What does OLIV do?

OLIV describes itself as voice typing for Thai mixed with English, for Macs. Hold the push-to-talk key (Right Option by default), speak, and release. The transcript is pasted at the cursor in whatever app has focus. The problem OLIV solves, in its README: Thai speech-to-text usually spells English technical words in Thai script, so a sentence about a build failing on a dependency version comes out in Thai letters. OLIV turns the English words back into English and leaves the Thai alone.

The pipeline runs on the Mac. Typhoon Whisper Turbo, a Thai fine-tune of Whisper, transcribes on the GPU through MLX. A Rust core applies a dictionary, sound-alike matching, Thai number and ๆ formatting, and your own Dictionary. A small on-device language model, Gemma E2B, restores English words and leaves the rest as spoken. The result is pasted with Cmd+V. A verbatim mode skips cleanup in terminals and password managers.

Key facts

• Licence: MIT.
• Speech model: Typhoon Whisper Turbo, by SCB 10X, a Thai fine-tune of Whisper, converted to MLX.
• Cleanup model: Gemma E2B 4-bit, through MLX. Total first-run download is about 5 GB.
• Platforms: Apple Silicon Mac (M1 or newer) with macOS 14 Sonoma or later. The signed, notarised DMG is 27 MB.
• Latency: median 1.0 second from releasing the key to text, measured on an M4 Pro by the README.
• Accuracy on the maintainer's held-out set: about 95 percent meaning match on 40 clips; about 98 percent on 194 everyday clips; about 97 percent on 30 confirmation clips.
• The README says these clips were recorded by one speaker (the developer), and "meaning match" is LaBSE similarity of at least 0.80, not word-for-word.
• Three run modes: on your Mac by default, your own OLIV server on a Linux GPU box, or Groq for speech-to-text with optional Gemma cleanup on the Mac.
• Features: push-to-talk or toggle, a Dictionary with CSV import and export, echo cancellation for playback, a 10-item recent-dictations menu kept in memory only, Thai number and date formatting.

What does it replace, and where does it fall short?

OLIV replaces Wispr Flow and Superwhisper for Thai speakers who mix English technical words into their speech. It is on the open-source alternatives to Wispr Flow and Superwhisper pages.

Where it falls short: Apple Silicon only, so Intel Macs and other operating systems are not supported. The accuracy figures were measured by one speaker, so other voices, accents and noisy rooms are untested. On the maintainer's held-out set, Whisper large-v3 transcribes slightly better before cleanup, and the README says so. Names and numbers can be wrong even when a sentence counts as a meaning match.

How does OLIV run?

Download OLIV.dmg from the releases page, drag OLIV to Applications, and grant Microphone and Accessibility access. On first run, OLIV downloads the Typhoon Whisper Turbo and Gemma E2B models from Hugging Face (about 5 GB once). Hold Right Option, speak, release.

Two other engines are optional and off by default: a self-hosted OLIV server on a Linux box with an NVIDIA GPU, reached by API key, with the Dictionary synced across devices; and Groq for speech-to-text in the cloud, with Gemma cleanup still on the Mac. Audio leaves the Mac only while a cloud engine is selected.

Who is OLIV for?

A Thai speaker on an Apple Silicon Mac who wants to dictate English technical terms correctly inside Thai sentences, without a subscription and without audio leaving the device. For a general English dictation app with the same push-and-paste shape, see Handy. For a hosted product, Wispr Flow or Superwhisper is the paid tool.

What limits does the README state?

• Apple Silicon only; Intel Macs are not supported.
• Accuracy figures were measured by one speaker, so other voices, accents and noisy rooms are untested.
• Names and numbers can be wrong even when the sentence counts as a meaning match.
• The raw Typhoon speech model is slightly less accurate than Whisper large-v3 before cleanup; OLIV wins on the full pipeline.
• Cloud engines (Groq, OLIV API) send audio off the device while selected.
• Role-play voice design and other TTS features are not part of OLIV; it is dictation only.

Questions people ask

Is OLIV open source?

Yes. The LICENSE file at the repository root is MIT, an OSI-approved licence. The README says OLIV is free, MIT-licensed, with no account and no telemetry. The downloaded models (Typhoon Whisper Turbo and Gemma E2B) are open-weight models hosted on Hugging Face, each with its own licence recorded on its model card.

Does OLIV work offline?

Yes, with the default on-device engines. After a one-time model download of about 5 GB, dictation works without internet and your audio is never uploaded. Cloud engines (Groq and OLIV API) are off until you turn one on, add a key and select it in the dictation engine panel.

How accurate is OLIV for Thai and English?

On the maintainer's held-out set of 40 clips, meaning match is about 95 percent, where meaning match is LaBSE semantic similarity of at least 0.80. On 194 everyday clips used for tuning, it is about 98 percent. Accuracy drops to 60 percent without the Gemma cleanup step. All clips were recorded by one speaker on 28 October 2026, so other voices and noisy rooms are untested.

How does OLIV compare with Wispr Flow?

OLIV is free, open source and focused on Thai mixed with English, running on the Mac by default. On the maintainer's 264-clip Thai-English set, OLIV led Wispr Flow on overall meaning match (97 percent against 93) and returned text sooner (median 1.0 second against 1.4), with Wispr ahead on heavy jargon (100 against 98). One speaker, measured 28 October 2026. The open-source alternatives to Wispr Flow page sets it beside Handy, VoiceInk and others.

Can OLIV type my own technical names correctly?

Yes, through the Dictionary. You add names, acronyms and jargon, star the ones that matter most, and OLIV spells them your way even when the speech model gets a Thai name slightly wrong. From 0.2.2 (October 2026) the Dictionary also fixes near-miss Thai spellings and lets an English entry replace its Thai spelling. CSV import and export is supported, and the OLIV API syncs the Dictionary across devices.

Sources

• OLIV README and LICENSE: github.com/chayapats/oliv, read 2026-11-02.
• OLIV benchmark: chayapats.github.io/oliv, referenced by the README, 2026-11-02.