Skip to content
moshpit

Docs · Topic

Voice input

Talking to a coding agent from a phone is often faster than thumb-typing a paragraph. Moshpit transcribes speech entirely on your iPhone or iPad—audio never leaves your device.

Choose your engine under Settings › Voice Input › Recognition:

Engine Best for Download
Local model — Qwen3-ASR (0.6B) Chinese mixed with English technical terms, flags, and identifiers in the same sentence, biased with hot words from the screen. One-time download under Model
Local model — Whisper Multilingual dictation across languages supported by WhisperKit. One-time download under Model
Apple (built in) Instant single-language dictation (set under Language) with zero extra storage. Built into iOS
  1. Open Settings › Voice Input › Model. Each row shows its download size before you tap anything.
  2. Tap Download beside the model you want (you can Pause and Resume at any time).
  3. Once downloaded, tap a row to mark it In use, or tap ⋯ beside it to remove the model and reclaim storage. Downloaded models live in Application Support and are excluded from iCloud backups.

When you dictate with Qwen3-ASR, Moshpit collects visible file names, the current git branch, and the active agent’s name from the pane on screen and passes them as hot words so project-specific names land spelled right.

  1. Tap the microphone chip on the key bar to start recording.
  2. Speak your prompt. Instead of streaming half-finished words straight into a live shell, Moshpit shows the transcription inside a preview card above the bar (along with the active engine and language).
  3. Fix any word if needed, then tap Insert to send the text cleanly into the pane via bracketed paste.