VoxBee has always been about keeping your voice on your machine. As of v0.6.3, that promise extends one step further into the pipeline: the auto-format pass that adds punctuation and capitalization to your dictation can now run entirely on-device via Apple Foundation Models — no network, no cloud LLM, no API key.
What It Does
When auto-format is enabled, VoxBee takes the raw transcription right before injecting it into your focused app and runs a quick Foundation Models pass that:
- Adds sentence-ending punctuation
- Capitalizes proper nouns and sentence starts
- Detects list cues ("first, second, third") and formats them as bullets where appropriate
The model runs locally on Apple Silicon. Your audio is already gone by this point — the model only sees the post-transcription text — and even that never leaves your Mac.
Requirements
- macOS 26 with Apple Intelligence enabled
- A device that supports Apple Intelligence (Apple Silicon required)
- Apple Intelligence's model files downloaded and ready
If any of these aren't met, the toggle stays available but shows you the specific reason the model can't activate (and falls back to your raw transcription).
Why Local?
You could absolutely do this with a cloud LLM. ChatGPT, Claude, or Gemini will happily punctuate a sentence. The reason to do it on-device:
- Privacy — even the cleaned-up text never leaves your machine
- Latency — no round trip; the format pass completes in under a second on most hardware
- Cost — zero per-dictation API charge
- Offline — works on a plane or in a dead zone
The Safety Net
On-device language models can occasionally do weird things — refuse to respond, return ALL CAPS output, or hang. Auto-format is engineered to never hold your dictation hostage:
- Hard 3-second timeout for live dictation (10 seconds for note dictation)
- All-caps output is rejected and the raw transcription is used instead
- Excessive length drift (output longer than 1.8x or shorter than 0.5x the input) is rejected
- Any error path silently falls back to the raw transcript — the user-visible behavior never breaks
This means auto-format is safe to enable: in the worst case, you get the same output you would have gotten without it.
When to Enable It
If you dictate long-form prose — emails, PR descriptions, blog drafts, journal entries — auto-format is almost always a win. If you dictate into structured fields where you've trained yourself to say "comma" and "period," you may prefer the existing behavior and want to leave it off.
Getting Started
Open Settings → Speech in VoxBee on macOS 26 and toggle "Auto-format dictation." Download VoxBee if you haven't yet — free 14-day trial, no account needed.