FireRedASR Large: The Best Offline Mandarin Dictation Model for Mac
For Mandarin-primary speakers who need dictation they can trust with professional content — documents, legal notes, medical records, correspondence — the question isn't whether to use a local model. It's which local model has the best Mandarin accuracy. FireRedASR Large is the answer.
3.18% character error rate
The FireRed team built FireRedASR using an attention encoder-decoder (AED) architecture with 1.1 billion parameters, trained on large-scale Mandarin speech data. On standard benchmarks, it achieves a 3.18% character error rate (CER) for Mandarin — state of the art for any offline, locally-runnable model.
To put that in context: a 3% CER on Mandarin means roughly 3 characters in every 100 are incorrect. In normal dictation of continuous speech, you'll see this translate to very few corrections needed per paragraph. For names, technical terms, or specialized vocabulary, accuracy can vary — but on general Mandarin speech, FireRedASR Large is as good as offline gets.
Mandarin, dialects, and English code-switching
FireRedASR handles more than standard Putonghua. It was trained on diverse Chinese speech data including dialectal variation, which means it handles regional accents better than models trained purely on standard Mandarin. It also supports code-switching between Mandarin and English — a common pattern in business, academic, and professional contexts where English technical terms appear in otherwise Mandarin speech.
If you regularly say things like “这个 deadline 是下周 五” or mix in English product names and acronyms, the model handles those transitions without requiring configuration.
The tradeoff: size and speed
FireRedASR Large is 1.7 GB — the largest model in Resonant. It also runs somewhat slower than SenseVoice Small given its size and autoregressive decoding. On Apple Silicon, it's perfectly usable for dictation; the pause between speaking and text appearing is longer than with lighter models, but not disruptive.
If you want faster Mandarin transcription at the cost of some accuracy, SenseVoice Small (226 MB, RTF ~0.10) is the alternative. For casual or high-volume dictation where speed matters more than perfection, SenseVoice is practical. For professional content where every character counts, FireRedASR Large is worth the extra time and space.
How to enable it
Open Settings → Transcription in Resonant and select “FireRedASR Large”. Allow time for the 1.7 GB download — Resonant runs normally with your current model while it downloads. Once ready, the model switches on your next dictation.
Everything stays on your Mac. Your Mandarin audio is processed locally on Apple Silicon and discarded immediately after transcription. Nothing is sent to any server.
Download Resonant and try FireRedASR for Mandarin dictation.