v0.2.2 (beta)

Transcripts now read like a conversation, not a stream of fragments

• One block per speaker turn — continuous speech from the same person now renders as a single readable block instead of splitting at every brief pause. A new block still starts when someone else speaks or after a long silence, so timestamps stay useful for jumping around.
• Stray punctuation cleaned up — a lone mark from the recognizer no longer opens a block of its own (". I am very tired") or doubles up ("day..") — it lands at the end of the sentence it closes.
• Long monologues summarize reliably — an uninterrupted stretch of talk can no longer overflow the on-device model and silently vanish from the summary and action items.
• Sharper speaker naming — the post-meeting "who is this?" clips and sample quotes are anchored to exact utterance timings, so they reflect a speaker's real speech.
• Auto-stop summaries stay clean — words captured in the few seconds after a call ends can no longer sneak into the summary.

Full changelog: v0.2.1...v0.2.2