VoxStation
A local AI studio for voice, music and video: dubbing, voices, full-cast audiobooks, songs, sound effects, translation and live captions — all on your computer.
Overview
CONTEXTVoxStation is a desktop studio for everything spoken and played, and nothing in it leaves the computer — no upload, no cloud bill. Eight workspaces share one engine and one look: Radio — endless songs written and sung in the mood you pick. It learns from what you like and skip, and can make hours of music ahead. For a café, a bar or a shop it is a stream of new music that never runs out, made on the venue's own computer and cleared for commercial use, with no royalties to pay. Music — a finished song from your own lyrics, sung word for word with a full band and a polished mix, or from a tune you hum into the microphone — the app hears its notes, writes the words and adds a band — or built from blocks (style, voice, instruments, parts). Redo any part; the rest stays as it was. Sounds — sound effects and game music from a description: several takes, trimmed, levelled and ready for an engine or an editor. Dubbing & Editing — transcribe, translate and re-voice a video in another language, timed to the original. Its montage editor makes stories for TikTok, Instagram and every feed: dozens of templates for promoting a product or marking an occasion, a large set of fonts, stickers and moving backgrounds, animation with keyframes, and export in any format from 9:16 to 4K. Voices — characters you can speak new lines as: a voice designed from a description of how it should sound, or a stock voice. Audiobooks — a manuscript in, a finished book out. A local model finds who says each line and casts a voice for every character; out come M4B with chapter marks, MP3 and a video. Translate — books, documents, subtitles and quick text, translated by a local model into over a hundred languages. Conversation — live captions for both sides of a call or a lesson, each line translated, with hints on what to say next. The interface speaks 49 languages, right-to-left ones included, and is built to WCAG 2.2: keyboard reach throughout, high contrast, larger text, reduced motion and reading fonts. The app is being prepared for Windows, macOS and Linux, at a fair price within everyone's reach.
The challenge
PROBLEMMake a dozen heavy open models — speech recognition, several voice engines, a music model, a sound model and a 9-billion-parameter language model — feel like one calm app on an ordinary gaming PC or a Mac. They do not fit on a 12 GB graphics card together, and they want different Python and CUDA builds. Each had to be loadable, unloadable and replaceable without the person ever opening a terminal — and a prompt typed in Ukrainian or Japanese still had to reach models trained on English.
Process
METHODOne engine, many models
A Python engine runs each heavy model in its own worker and environment; the desktop shell starts it and talks to it over a local port only.
Memory you can see
Models load when a workspace needs them. A status-bar widget shows each one's own share of the graphics card and memory, why it is kept, and unloads it — or stops the radio and unloads — in one click.
Set up per machine
The installer stays small: on first run it builds the right environment for the machine — NVIDIA, CPU or Apple Silicon — and downloads only the models chosen.
Every language
49 interface languages with a layout test for each, right-to-left layouts, and any prompt put into English before it reaches a generator.
Accessible by design
Every screen built to WCAG 2.2: full keyboard reach and a visible focus ring, high contrast, text up to 150%, reduced motion, Atkinson Hyperlegible and OpenDyslexic, and an accessibility statement in the app.
Pipeline
PROCESSText, a video or a recording in → speech recognition (Whisper) → the local language model reads, marks up, translates or writes → a designed or stock voice, the music model or the sound model makes the audio → it is aligned to the text and the picture, mixed and levelled → out as M4B, MP3, WAV, video or subtitles, into folders the person can see.
Technical notes
DEPTHA Tauri (Rust) desktop shell around a FastAPI engine in Python and a Next.js, React and TypeScript interface served as a static export. Speech recognition with Whisper. The local language model is Qwen 3.5 9B through Ollama: it marks up books for a cast, translates, writes lyrics and puts any prompt into English. Voices are designed from a description with Qwen3-TTS and read by Qwen3-TTS, VoxCPM2, HolosTTS and Chatterbox; Piper and Supertonic give stock voices; songs from ACE-Step 1.5; sound effects from MOSS-SoundEffect. Each heavy model runs in a worker process of its own with its own environment, built for the machine by uv. The engine knows what each one holds on the card and in memory, measured per process, and lets the person unload it. The interface is reachable by keyboard throughout, with high-contrast and larger-text modes, themes, and one shared waveform design for every player.
Outcome
RESULTIn development for Windows, macOS and Linux. Everything shown here runs on one machine with a 12 GB graphics card. It grew from the same roots as PolyBuilder, and the two share their look.
Gallery
MEDIARadio — A radio that writes its own songs
Radio for business — Endless music for your café, bar or shop
Music — Songs from blocks, words or a hum
Music · Hum it — Hum a tune, get a whole song
Music · Your words — Your own words, sung as a finished song
Sounds — Sound effects and game music from a description
Dubbing & Editing — A video, spoken again in another language
Stories & montage — Stories for TikTok, Instagram and every feed
Stories & montage · A story for every occasion in one press
Stories & montage · 52 text styles, each with its own look
Stories & montage · Over 1,000 animated stickers
Stories & montage · Animated pixel-art stickers
Stories & montage · Every emoji, ready for your video
Stories & montage · Moving backgrounds that dance to the music
Character voices — A voice for every character, made from words
Audiobooks — A manuscript in, a finished audiobook out
Audiobooks · Cast — Every character gets a voice of its own
Translate — Translation that stays on your computer
Conversation — Live captions and hints for any conversation
Models — Every model in memory, in plain sight
Accessibility — Made for everyone, to WCAG 2.2
First run — Tuned to this computer on first run
Set-up — A small installer that sets the machine up
Updates
CHANGELOG- 2026-10-14Eight workspaces — radio, music, sounds, dubbing and montage, voices, audiobooks, translation, live conversation; 49 interface languages; model memory control; Windows, macOS and Linux.
Working on something?
Open to frontend, full-stack and game-dev roles — hybrid near Cardiff or remote.