Qwen3-ASR, locally
Qwen3-ASR on your Mac,
as a dictation app.
Qwen3-ASR is Alibaba's open speech recognition model. It covers 52 languages and dialects, 22 of them Chinese dialects, and handles sentences that mix languages. Running it yourself means Python, MLX or a build of llama.cpp. TypeAloud packages it: download the app, pick the model, hold a key and talk.
Qwen3-ASR in TypeAloud, measured on an M4 with 24 GB
| 0.6B (default) | 1.7B (optional) | |
|---|---|---|
| Download | About 1 GB | About 2.52 GB |
| Transcribing a 6-second clip | 0.30 s | 0.71 s |
| Transcribing a 30-second clip | 0.96 s | 2.07 s |
| Model load | About 0.7 s | About 1.0 s |
| Peak memory | 1.39 GB | 2.89 GB |
Competitor details checked on 26 Oct 2026 against Qwen3-ASR model card. Prices and limits change; the official pages are authoritative.
How it runs
The model runs on your Mac's own chip, so recordings are never uploaded. Technically, TypeAloud runs 8-bit (Q8_0) builds of Qwen3-ASR through llama.cpp, using Metal on Apple Silicon. A worker is warmed while you speak, so most of the load time overlaps your sentence. On Windows and Linux the same models run on the CPU, with optional Vulkan on Windows; those builds are experimental.
The timings above exclude recording and insertion, and come from the TypeAloud maintainer's measurements on an M4.
Dictating for a long time
Qwen3-ASR works best on clips of a minute or less, so TypeAloud cuts long dictation at natural pauses into pieces of 20 to 75 seconds and transcribes each while you keep talking. When you release, the pieces are joined, without stray spaces between Chinese characters, and inserted once. TypeAloud does not import audio files.
Which size to use
Start with 0.6B. It is fast on any Apple Silicon Mac. TypeAloud records how quickly your Mac transcribes, and once it has enough samples it suggests 1.7B if a 15-second sentence would still come back within about 1.5 seconds. Whether 1.7B is more accurate on everyday dictation has not been measured, so treat it as an option, not an upgrade.
No language setting
TypeAloud uses Qwen3-ASR's automatic language detection, so you never choose a language. That is why mixed Chinese and English works without switching anything, and why it is a common answer to Whisper's trouble with mixed speech.
Common questions
- Do I need Python or Homebrew?
- No. TypeAloud is a signed Mac app. The model downloads from inside the app the first time you choose it.
- Does it need the internet after the download?
- No. Transcription runs on your Mac. See offline dictation for Mac.
- Can I use it from the command line?
- Not currently. TypeAloud is a dictation app: hold Ctrl+Shift, speak, and text appears at the cursor.
- Is it open source?
- Yes, TypeAloud is GPL-3.0. See open-source dictation for Mac.
TypeAloud: free dictation for Mac, transcribed on your Mac.
Hold Ctrl+Shift, speak, release, and the text appears at your cursor. No word limit, no account, open source.
Download for Mac