v1.3.2

Soniox settings gained "Context tags." You can use key-value pairs like domain / topic / role to tell the model your context — for example domain: software engineering, topic: web3 trading bot, role: backend engineer. The model treats these tags as priors during recognition, noticeably cutting misrecognition of technical terms and abbreviations. Every Soniox user is different, so these tags are fully customizable, up to 6 pairs, and take effect as soon as you save.

The vocabulary list is now a "chip tag" interface. What used to be a list of hand-typed correction entries now works like tagging — type to add, click × to remove — with every change staged locally and pushed to Soniox / the vocabulary list all at once when you hit "Save." That means you can make a dozen edits in one go without any single one triggering a mid-edit reconnect and connection jitter. Voilà also now ships with a built-in default correction — voila → Voilà — so you don't have to add it yourself.

The four provider settings pages (Soniox / ElevenLabs / Groq / Volcengine) got a complete visual redo. The auth section used to be a custom card sitting above a native macOS Form group — two completely mismatched visual styles, with the lower half even having its own separate scrollbar. Now the whole page is one unified scroll area with one card style throughout — consistent top to bottom, matching widths, and an opaque background on the save-button bar for better contrast. Browsing settings day to day no longer feels like two pages stitched together.

Fixed a hidden bug in Soniox's async mode: selecting stt-async-v4 could get a recording mid-way misjudged as a dead connection by the "first-token watchdog," forcing a reconnect that swallowed the whole audio segment. The root cause was that the watchdog only distinguished streaming vs. batch by provider, without accounting for Soniox being a single provider that spans both modes. Async-mode recording is stable now.