Qwen3 ASR: The Best Multilingual Offline Dictation Model for Mac
Resonant ships with two speech-to-text models. Parakeet TDT covers English and 24 other European languages with the lowest word error rate of any locally-runnable model. Qwen3 ASR is the other one — and if you dictate in more than one language, or in any language outside Parakeet's set, this is the model you want.
What Qwen3 ASR actually is
Qwen3 ASR 0.6B is an open-weight automatic speech recognition model from Alibaba's Qwen team. It uses a transformer encoder-decoder architecture, sized at roughly 0.6B parameters, and weighs in around 1.7 GB after compilation for CoreML. In Resonant, it runs entirely on Apple Neural Engine — no cloud, no audio leaving your Mac.
Unlike most ASR models, which are tuned to one language family at a time, Qwen3 ASR was built from the ground up for broad multilingual coverage. The trade-off is small: its accuracy on any single language is slightly lower than a specialist like Parakeet, but the spread is much wider and the multilingual behavior is materially better.
30+ languages, including the ones most local models miss
Qwen3 ASR covers more than 30 languages, including ones that are underserved by every other locally-runnable option:
- Mandarin Chinese (Simplified and Traditional)
- Cantonese
- Japanese
- Korean
- Vietnamese
- Thai
- Indonesian and Malay
- Arabic
- Hindi and other Indic languages
- Turkish, Hebrew, Persian
- All major European languages
If you dictate in any of these — and especially if you're used to either sending audio to the cloud or fighting with Whisper variants on Apple Silicon — Qwen3 ASR is the upgrade.
Code-switching: the feature most models can't do
Code-switching is when you change languages mid-sentence. A bilingual professional dropping a Mandarin idiom into an English email. A doctor in São Paulo using English drug names inside a Portuguese note. A French engineer narrating code in English.
Most ASR models freeze in this situation. They'll pick one language, stick with it, and phonetically mangle everything that doesn't fit. Qwen3 ASR was trained with code-switching as a first-class case — it follows the language as it shifts, inside a single utterance, without you needing to change a setting.
In Resonant, this means you can speak naturally. The model handles the mixing.
Who should use Qwen3 ASR
Pick Qwen3 ASR if any of these are true:
- Your primary dictation language is one Parakeet doesn't cover — Mandarin, Japanese, Korean, Arabic, Hindi, Vietnamese, Thai, etc.
- You routinely code-switch between two languages in the same message.
- You work across more than three languages in a typical week and want one model that handles them all.
- You want a fully local, open-weight alternative to cloud multilingual ASR services.
If you only dictate in English, or English plus one other European language, Parakeet is the better default — it's smaller, faster to load, and slightly more accurate inside its language set.
How to use it in Resonant
Both models ship pre-installed with Resonant, so there's no separate download. To switch, open Settings → Transcription and select “Qwen3 ASR 0.6B” from the model list. The change applies immediately.
You can move back and forth between Parakeet and Qwen3 ASR as your work changes. Some people keep Parakeet as the default and switch to Qwen3 ASR for sessions in another language. Others run Qwen3 ASR full-time because their work is multilingual by default.
Either way, audio stays on your Mac. The model runs on Apple Neural Engine, the transcript is generated locally, and the audio is discarded the moment transcription completes.
Download Resonant to try Qwen3 ASR on your Mac.