Changelog

Launch-readiness pass ahead of the first TestFlight→App Store push. All transcription still runs on-device; these changes harden the boundary between private data and the rest of the app.

• Privacy

  Opt-in PII redaction on every export format — phone numbers, emails, credit cards, SSNs, and IP addresses get replaced with category tokens.

• Privacy

  Persistent custom vocabulary so proper nouns you care about never get mis-transcribed a second time.

• New

  On-device named-entity detection with a reader-strip that jumps to the first mention of a person, place, or organization in the transcript.

• New

  Audio-event pills — [music], [applause], [laughter] now render as styled chips so non-speech moments are scannable.

• Polish

  Settings, Reader, and Export sheet all collapsed to a first-run-friendly footprint. Power-user knobs live behind Advanced disclosures.

Real-time mic transcription and a faster on-device engine. First build that routes through the full engine matrix (WhisperKit / Apple Speech / Moonshine) depending on your content.

• New

  Live transcription — tap the mic from Voice tab, speak, watch words land on screen in real time. Saves as a regular transcript when you're done.

• New

  Moonshine engine integration — 5× faster than Whisper for short clips, fewer hallucinations on music-heavy audio.

• New

  Snip suggester — AI picks 3-8 viral clip candidates per episode using density, speaker diversity, and quotable-phrase heuristics.

• Faster

  Library scrolling on 100+ transcript libraries stays 60fps — per-cell blur passes dropped from 30fps to 15fps, backdrop materials scoped to the active segment only.

On-device 270M-parameter language model wired up for the AI Companion and a new verification pass that catches Whisper hallucinations before they reach the reader.

• New

  AI Companion — an on-device chat partner that knows every transcript in your library, powered by FunctionGemma running via MLX.

• New

  Hallucination filter — second-pass verification flags 'thanks for watching' / 'subscribe' artifacts WhisperKit sometimes emits during silence.

• New

  Three transcription engines, one picker — WhisperKit, Apple Speech, Moonshine. The engine auto-selects based on your Quality Profile.