v2.5.0: A Full Voice Overhaul, 30 Previewable Voices, SaaS Calling Fixes, and a Reliability Batch
If you have ever been on a phone call with an agent that talked over you, repeated itself, or cut you off mid-sentence, you know exactly how quickly it breaks the illusion that you are talking to a competent colleague. v2.5.0 is largely about closing that gap. It is a big release, and the center of gravity is voice.
Your voice agent stops talking over you
The headline fix: calls were cutting the caller off and sometimes repeating the previous reply. We traced it to the anti-stall watchdog in our Gemini live bridge. That watchdog exists to nudge the model if it stalls, but it was firing during the window where you, the caller, were still speaking. The nudge forced a stale re-generation right over the top of you.
The watchdog now stays quiet while you are talking. On the ElevenLabs path we cleaned up four more bugs in the same pass: text getting swallowed on a phantom interrupt, the agent saying goodbye twice, onomatopoeia being read out literally, and a six to seven second wait for the first word. Both voice stacks were validated on real calls, not just in a test harness.
Thirty voices, one source of truth, with previews
Voice selection was quietly a mess behind the scenes. The phone (Twilio) path and the in-app voice call were reading from four separate, slowly diverging copies of the voice list. We collapsed those into one source of truth.
For you, that shows up as choice. Twilio went from eight voices to thirty, and every one has a preview button so you can actually hear it before you assign it, rather than picking blind from a name. The in-app voice call now uses that same picker with the same previews. And voices are filtered by the employee's gender, so a female employee offers female voices and a male employee offers male ones, on both screens.
If your account language is French, the previews play French samples for all thirty voices. Everyone else hears English. It is a small touch, but hearing a voice in your own language before you commit to it makes the choice feel real.
The Twilio Voice settings panel got a full redesign too, reorganized into clear sections: Number and voice, Conversation, Capabilities, and Personalization, with a compact modal picker. For now Twilio runs Gemini Live only, because the ElevenLabs model Twilio exposes is an older, expression-less one, and we would rather give you the good version than a flat one.
SaaS calling that tells you what it needs
On the SaaS side, calls run on your own billed Gemini key. The problem was that the product did not make that obvious, so calls would fail at the worst possible moment, when you tried to place one. Now there is an inline field to enter your key right where you need it, on the Start Call panel and in Twilio Voice settings, since SaaS has no separate API Keys menu.
Start Call and number assignment stay disabled, with a clear hint, until that key is set. So instead of a mysterious failure at call time, you get told up front what is missing. We also fixed a case where your owned Twilio numbers were invisible if no Gemini key was set (listing numbers only needs the Twilio credentials), and killed an infinite config request loop when expanding Twilio in SaaS.
Connectors stop tripping the rate limiter
Opening the Connectors panel used to fire off a burst of around twenty-four separate status requests, one per connector, which in the worst case tripped the rate limiter and threw 429 errors that could block your configuration. There is now a single aggregated status endpoint, so opening the panel makes one request for all the badges. Statuses also load the moment the panel opens, instead of waiting for you to expand each card.
Kanban, and the fixes that keep the lights on
The Dashboard Kanban now reuses the chat's board components, so the columns finally render side by side correctly (they were misaligned before) with the same rich cards: image grids, comments, badges. And if an employee is shared with you, you can now edit the board, creating, moving, updating, and deleting cards, not just the owner.
The rest is the unglamorous reliability work that matters more than it sounds. Fresh installs were broken because some SQL statements sat in the wrong order in our migration file and aborted the whole thing on a brand-new database; that is reordered and verified against a from-scratch schema. Agent containers now inherit their owner's timezone instead of defaulting to UTC, so anything time-sensitive fires at the right local hour. The Read-tool image preview shows again, survives a refresh, and now handles large images. Orphaned Reassign and Delete sheets in Manage Employees now close with their parent.
On the surface: markdown now renders live while the assistant streams, so bold, italic, and inline code appear in real time instead of flashing as raw tokens. The collapse threshold went from 300 to 1024 words, so a normal long technical answer is not hidden behind an Expand button. And usage statistics now spell out their periods clearly: week means the calendar week, month means the current calendar month, both in your timezone.
None of these fixes are flashy on their own. Together they are the difference between an AI employee that feels like a demo and one that feels like it actually works, especially the moment it picks up the phone.
Want to test the most advanced AI employees? Try it here: https://Geta.Team
