Last Word No Longer Cut Off — Streaming Edition
Real-time streaming with Soniox and AssemblyAI now reliably captures the final word, even when you release the hotkey the instant you finish speaking. Previously, the streaming pipeline closed the WebSocket too eagerly on key release, dropping any audio that hadn't been delivered yet. The connection now stays open for a brief tail period so the trailing audio reaches the recogniser before the session closes — matching the behaviour batch transcription has had since 1.2.
Faster Command Mode
Command mode (select text → hold hotkey → speak instruction) now finalises in about half the time. The session-close wait had quietly grown to 4 seconds across recent builds; we've taken it back down to 2 seconds, matching v1.3.7. Most short commands now complete in roughly a second from when you release the hotkey.
Smarter Command Mode
The AI behind command mode is now llama-3.3-70b-versatile (was llama-3.1-8b-instant). The smaller model was fast but wrote off-script — paraphrasing instead of transforming, echoing the prompt as preamble, dropping content. With 70B in the loop, "make this professional" produces a genuine rewrite, "translate this to Serbian" actually translates, and "fix grammar" leaves your meaning alone. The cost is roughly 200–300 ms per command, well worth it.
Bug Fix: Cursor Check in Browsers
The "cursor not in a text field" check was too lenient on web pages — any focused element that responded to accessibility queries was treated as editable, even when you weren't in an input. The check now requires positive evidence: a non-empty selection, an explicit editable hint, or a focused descendant that's actually a text input. Fewer false positives in Chrome and Safari.