v0.3.0 — Faster, more accurate verbatim transcription (turbo)

Highlights

Verbatim transcription switched to the turbo model with a per-chunk filler prompt. A/B testing across three real episodes showed the new method beats the previous large-v3 approach on every axis:

• More fillers captured (e.g., 60 → 88 on a 24-minute episode; independently verified against the audio)
• Fewer hallucinations — large-v3 produced repetition loops on long quiet sections; turbo with 30-second fresh-context chunks does not. A new time-density guard catches rare boundary hallucinations and rescues them with a local large-v3 re-recognition
• 2x+ faster — a 25-minute WAV now completes in ~1.5 minutes end-to-end on an M4 Pro (was ~4 minutes)
• Cleaner transcripts — turbo's natural orthography makes the .sc.txt read like ordinary text, which makes text-based editing genuinely pleasant

Other changes

• Filler dictionary: added えーっと (was in the prompt but missing from the dictionary)
• Edit stats now show seconds removed for filler cuts and deleted lines, e.g. Fillers: removed 33 of 60, -14.9s
• Logic export annotates the Duration line to clarify that cuts remain as timeline gaps in FCPXML
• README / README_ja / llms.txt updated to match

Full diff: v0.2.0...v0.3.0

v0.2.0

First tagged release of stefnceorf — a local, free CLI that cleans up speech recordings for podcast production.

Highlights

• One-command pipeline: sc input.wav transcribes, trims silences, removes filler words, and exports FCPXML for Logic Pro (File > Import > Final Cut Pro XML).
• Silence trimming: long pauses are shortened to a natural length (--gap-threshold / --gap-max), both in rendered WAV and on the Logic timeline.
• Filler removal: suggested fillers (えー, あの, etc.) are cut only when acoustically safe — cut points snap to waveform troughs, and region edges snap inward to quiet points to prevent clicks.
• Optional text-based editing: edit the generated .sc.txt (delete/reorder lines, remove words, keep/discard 〔〕 filler suggestions) and re-render with sc render / sc logic.
• Hallucination containment: suspicious transcription spans are re-recognized; unrecoverable spans are kept as explicit "unrecognized" regions instead of silently dropping audio.
• Local & free: powered by mlx-whisper on Apple Silicon. Japanese-first, English supported.

Commands

Command	Purpose
sc input.wav (= sc auto)	Full pipeline: WAV → FCPXML for Logic Pro
sc transcribe	WAV → .sc.txt / .sc.json
sc render	Edited .sc.txt → .edited.wav
sc logic	Edited .sc.txt → .logic.fcpxml

373 tests passing.