Publishing from voice

Record on your phone, run one command, get a Sanity draft ready to review. Transcription via Cloudflare Whisper.

The fastest way to get a thought out of your head and into the drafts queue: record it as audio, run one command. No typing, no Studio, no git.

What happens

1. You record audio anywhere — Voice Memos on iOS, Recorder on Android, QuickTime on Mac, arecord on Linux
2. You run bun run publish-audio path/to/file.m4a
3. The CLI ships the audio to Cloudflare Workers AI (Whisper) for transcription
4. It structures the transcript — first sentence becomes the title, next few become the description + TL;DR, the rest is grouped into paragraph blocks
5. A draft post lands in Sanity with a Studio edit link printed to your terminal
6. You review, polish, and hit Publish when it’s ready

Command

bun run publish-audio ~/Downloads/voice-note.m4a

Output looks like this:

📼 voice-note.m4a · 857KB · m4a
🎙  Transcribing via Cloudflare Whisper...
   ✓ 55 words in 2.8s

💾 Creating DRAFT post: "This is a voice note from the workshop"

✓ Done
  _id:    e0SvlUp2bPBVJv2KpcLRei
  slug:   this-is-a-voice-note-from-the-workshop
  status: DRAFT — review in Studio

  Studio edit link:
  https://ded-blog.sanity.studio/desk/post;e0SvlUp2bPBVJv2KpcLRei

Click the Studio link, do your editing, hit Publish. The webhook rebuilds the site.

Flags

--title="Override the auto-title"     # Skip title extraction
--category="AI Agents"                 # Default: Engineering
--live                                 # Publish immediately (default: draft)
--raw                                  # Skip structuring, use raw transcript
--dry                                  # Print what would be sent, don't hit API

Formats supported

Anything Whisper accepts: .m4a, .mp3, .wav, .mp4, .webm, .ogg, .flac. Max 25MB per file. If your recording is longer, split it — most voice memos under ~20 min will fit.

Environment

The audio CLI reads:

SANITY_PROJECT_ID=3onlytdh
SANITY_WRITE_TOKEN=sk...
CLOUDFLARE_ACCOUNT_ID=d810128e93e1fd0f94b4948fd3edadfb
CLOUDFLARE_API_TOKEN=cf-token-with-workers-ai-scope

The Cloudflare token needs Workers AI Read permission (or Account Read + Workers AI). Everything else uses tokens you already have configured for the text publish flow.

Recommended capture apps

• iOS: Voice Memos is fine. Share → save to Files → run the CLI on your Mac against that path
• Android: Google Recorder (does live transcription too — but Whisper is more accurate)
• Mac: QuickTime, or sox -d 60 out.wav in the terminal for scripted captures

Why draft-by-default

Whisper is good but not perfect. Proper nouns get mangled (Sanity → “insanity” is a real one), homophones slip through, and the paragraph grouping is heuristic-only. Landing as a draft gives you 30 seconds in the Studio to fix the two or three things Whisper missed before it goes public.

If you’re recording something short and low-stakes where you don’t care about polish, add --live to skip the draft step.

Limits and edges

• 25MB max per file — Whisper API cap. About 20 minutes at reasonable quality.
• English works best. Whisper supports 99 languages but I’ve only tuned the structuring heuristics for English.
• Filler words survive — “so,” “um,” “you know” all transcribe. Edit them out in the Studio, or dictate more deliberately.
• Homophones need review — always look at the first pass before publishing.
• The paragraph split is dumb (every 4 sentences). If you want cleaner paragraphs, use --raw and structure it yourself in the Studio.

When to use this vs the text CLI

• Voice CLI (publish-audio) — you’re walking, driving, on a plane, or the thought is more emotional than technical
• Text CLI (publish) — you’re at a keyboard, drafting alongside code, or the post has technical structure (code fences, tables) that voice won’t produce

Both write to the same Sanity dataset, both hit the same webhook, both produce the same downstream flow.