Voxta v1.10.2: Memory Books, New Voices & Coherent Pictures (Patreon)

Hello everyone,

Since v1.10.0 we have shipped two updates, and this post covers both. They were quality releases, so a good part of what follows is fixes, but three things did grow. Memory books are now something you curate: entries can be marked as validated so a merge cannot touch them, reconciliation runs in small groups you can watch and stop, and documents are read a chunk at a time so a long one goes in whole. Voices gained a new text to speech engine that streams, a new streaming transcription backend, and a language picker that only offers what the engine can actually say. And generated pictures are now tied to the character they belong to, which is what keeps them coherent in a chat with several characters and in /imagine.

There is a fair amount here, so let's get started.

🧠 Memory Books

Recall used to treat every entry the same, whether you wrote it or the extraction did. Now you can say which ones you trust.

Memory entries can be marked as validated, and unvalidated ones recall lower, recall last, and are shown as unconfirmed

A merge skips validated entries, so curating a book is not undone by the next pass

Merge reconciles a book in small groups, one prompt each, with progress and a stop button

Documents are extracted a chunk at a time, cut at markdown headings, then paragraphs, then sentences, so large documents go through

Extract takes text and markdown files, or pasted text, straight into the editor

Extraction returns its keywords instead of dropping them

Memory books have Merge and Extract tabs, with sensitivity, progress, live results, and a validated or keyword filter

Closing a chat no longer runs memory passes the character has turned off

🎙️ Voices & Speech

VibeVoice text to speech, which streams, trims the silence around a voice sample, and caps its length

Deepgram Flux speech to text

Speech services report the languages they support, and the language is only offered where the engine uses it

A character can use a separate voice language, set in the Voice tab, independent of its culture

A line that is only a bracketed note is no longer spoken, instead of producing a clip that can never play

Updated SenseVoice and WhisperLive

🎨 Pictures

Generated pictures are tied to a specific character, so a chat with several characters keeps each picture with its subject, in /imagine as well

Regenerating a picture of a character carries the character's appearance on the second attempt too

Write an image prompt for a scenario thumbnail

Chat attachments are stored at the assistant's resolution

The attachment remove button sits on the thumbnail, and clicking the thumbnail opens the image viewer

A warning when a scenario image is missing

Download a thumbnail from the Assets tab

Better backdrop behind the image generating placeholder

🛠️ Tools & Agents

Claude Code and Codex tool calls now ask for approval

Allowing a tool Always is saved per chat

Up to 128 tool call rounds, and the same tool can be called more than once

Codex receives images, and both Claude Code and Codex can be used for computer vision

Computer vision images are scaled down before being sent

Diagnostics show the images that went with an inference

Windows SDK: the PowerShell tool strips the color escape sequences

📁 Files & Folders

A file browser when picking folders or files on disk, instead of typing the path

Administrators can browse the server's files and folders

Open a server folder in the file explorer on the server's own desktop

🔒 Shared Servers

If you run Voxta for yourself, nothing here changes. On a server with more than one account, the settings that reach the machine itself are now administrator-only.

Paths, API URL overrides and model downloads are administrator-only

Model repository browsing and downloads are administrator-only, and a model reference cannot point outside the models folder

ComfyUI, Oobabooga, OpenAI Compatible, TTS HTTP API, Claude Code and Codex are administrator-only modules

📦 Characters, Scenarios & Packages

Characters and scenarios can refer to private lore books, so an imported character's book is not listed globally and is deleted with the character

Packages update the explicit level and thumbnail when a contained resource is saved

Tags and a favorite toggle on a scenario's Overview tab

Running chats are marked on their card

⚙️ Models & Backends

OpenAI Compatible lists the models a server offers, where it offers them

Updated ExLlamaV3 to 1.4.9

Lower memory use from diagnostics

🖥️ Voxta UI

Talk view captions follow the message being spoken, and the spoken line takes the configured accent color

The talk waveform is drawn at screen resolution, in the character's color, and re-centers when subtitles or the input are hidden

A character's emote color lasts past the sentence, until the next emote or the end of the chat

A stage avatar with no transparency is framed as a portrait instead of sitting as a hard rectangle among the cutouts

Mini select menus size to their longest option

Settings forms stay responsive while dragging sliders

🔧 Key Fixes

Both updates were quality releases, so this list is longer than usual. The ones most likely to have reached you:

Tool results reported only their first line, so anything longer came back truncated

OpenAI Compatible services such as LMStudio tried to continue an event and returned empty replies

Whisper's GPU choice leaked into every later process, which sent local image generation to the wrong card

The diffusers VRAM cap applied to device 0 regardless of the card in use

ElevenLabs dropped request stitching

Unspoken reply lines appeared before their voice, and a silent line stayed pending until the reply had finished playing

A reply that is only a bracketed note rendered as an empty bubble

Audio preloading

Claude Code did not work in the Scenario Assistant

Message visibility for absent characters did not work with characters disabled at chat start

Scenario roles were not correctly linked to characters

Failed image generations looked like they were still loading, and failed inferences showed a NaN duration

The insecure page warning repeated on every chat

Service test transcripts landed in the chat input

The TTS Playground showed a timeout while the voice engine was still loading

🆓 Voxta Lite

New to Voxta? Lite is free and runs on Voxta Cloud, so you can create an account and start chatting without any local setup. If you like it, a subscription gets you the full version, with a large library of integrations, custom modules, and the option to run everything on your own hardware.

Thank you all, as always, for the continued support and feedback. Whether you are curating a memory book you can finally trust, giving a character a voice in a language of its own, or watching pictures stay with the character they belong to, we hope you enjoy this release.

Voxta is available to subscribers at: https://portal.voxta.ai/