omniASR 300M: Offline Dictation for 1,600+ Languages on Mac

Most speech recognition models support a handful of languages. Whisper, which is extraordinarily broad by modern standards, covers 99. Meta's omniASR covers over 1,600. If you speak a language that no other model in Resonant handles — or you work with audio in languages that don't have well-resourced offline models anywhere — omniASR is where you go.

What 1,600 languages actually means

There are approximately 7,000 living languages. Most speech recognition research has focused on the 20 or so that dominate global internet usage. The result is that speakers of the other 6,980 have had essentially no offline dictation options.

Meta built omniASR as part of their work on universal language technology, applying CTC architecture to cover the long tail of human language. The 300M parameter model was trained on audio from over 1,600 languages, including many African languages, indigenous languages of the Americas, Pacific Island languages, and other low-resource languages that have never had any practical offline ASR option.

The 348 MB model covers Swahili, Amharic, Yoruba, Hausa, Tagalog, and hundreds of other languages that no other model in Resonant addresses. For speakers of these languages, omniASR may be the only locally-runnable option that exists.

Accuracy expectations

For well-resourced languages like English, Spanish, French, or Mandarin, omniASR's accuracy is good but generally lower than the dedicated models. Parakeet is better for English and European languages. SenseVoice and FireRedASR are better for East Asian languages. GigaAM v2 is better for Russian.

For lower-resource languages, the comparison is different: often there is no dedicated model to compare against. omniASR's accuracy on a language like Tigrinya or Kinyarwanda may be imperfect, but it represents a baseline that previously required cloud infrastructure to access at all.

If your primary language is covered by another model in Resonant, use that. omniASR is most valuable as the option for everything else.

How to enable it

Open Resonant Settings → Transcription and select “omniASR 300M”. The 348 MB download is quick. No language configuration is required — omniASR auto-detects the language from the audio.

All processing happens on your Mac. Your audio in any of those 1,600+ languages stays entirely on your machine and is processed locally by Apple Silicon. No network connection is needed after download.

Download Resonant and try omniASR for your language.