Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableTG: @neurogen_news (Neurogen)→ original

Neurogen Releases an Offline Voice Studio: 16 Voices in One Dialogue and 8 GB of VRAM

Cloud voice-over services charge for every failed take and send your scripts and voices to someone else's servers. Neurogen has released a Windows app built on the Fish Audio Drama 3 model that runs on your own computer. It clones a voice from 5–15 seconds of audio and handles dialogue for up to 16 characters, and an NVIDIA GPU with 8 GB of video memory (VRAM) is enough. You need the internet only to download the models, and takes cost nothing.

AI-processed from TG: @neurogen_news (Neurogen); edited by Hamidun News
Neurogen Releases an Offline Voice Studio: 16 Voices in One Dialogue and 8 GB of VRAM
Source: TG: @neurogen_news (Neurogen). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @neurogen_news (Neurogen) ↗
Editor-in-Chief
Emotions are set with tags in the textwhisper, laughter, sigh, and pauses
Key metric #2Export to FLAC for editing; projects are saved as .drama3 files
Key metric #3Requires Windows 10/11, an NVIDIA Turing+ GPU, and 32–64 GB of RAM

Imagine you are making a game or an online course and need to voice hundreds of lines. In cloud speech synthesis services, every failed take eats up credits, and your scripts and other people's voices go to third-party servers. The Neurogen team has released a portable voice studio that runs entirely on your computer.

It is built on the Fish Audio Drama 3 speech synthesis model (it reads text aloud in a voice taken from a short sample), and a built-in ComfyUI engine handles fast startup. The studio assembles a dialogue of up to 16 characters in a single script, and an NVIDIA GPU with 8 GB of video memory (VRAM) is enough to run it. If you have a lot of voice-over work, this is a way to stop paying for every take and to keep your voices out of the cloud.

How the studio works

Python, PyTorch, and ComfyUI are already bundled inside the app, so you don't need to install anything manually. The engine starts automatically or manually, whichever you prefer. The models are downloaded separately from within the program: downloads resume after an interruption, and the files are verified. After that, the studio works without the internet, and your texts and voice samples are not sent anywhere.

The whole process lives in one window: prepare a voice, write a script, try out the delivery, listen to the variants, save the project, and export the audio for editing. For editing, there is FLAC export. For those used to working with nodes, the authors provide a workflow for a compatible ComfyUI.

Voices, dialogues, and emotions

Cloning is simple. You upload a recording, select 5–15 seconds of clean speech in it, and use that voice for new text. The voice, together with its transcript, can be saved to a library, so you can return to it in your next video or for one character's lines. The authors remind users to use only recordings they have the rights to.

For a scene with several characters, you assign each one its own sample, and up to 16 participants are combined into a single script. You can also start with an ordinary monologue. You change the delivery directly in the text: select a phrase and add a whisper, laughter, a sigh, or a cue for a cheerful intonation, and set pauses with tags.

You can also find a voice without a source recording. The model has starter options, and random selection helps you discover new ones. You can lock in a result you like and save it to the library. However, what gets saved is the seed of the successful generation (a number the result depends on), so the timbre may change on a different text. For a recurring character, cloning the voice from a recording is more reliable.

A separate sample editor shows the waveform and offers playback, precise boundaries, and trimming. Demucs can separate speech from music, and the original file is kept for comparison. The sample's text is entered manually or recognized during generation via Qwen3-ASR.

Projects and hardware requirements

A project is saved in a .drama3 file, which holds the script, the settings, and the voice samples themselves. The "Return to studio" button restores the settings of the selected result, and autosave and an archive of drafts let you pick up the work later. For long texts, there is automatic splitting, reading the whole text or line by line, separate takes, and shortening of long pauses. Advanced mode unlocks model selection, seed, and generation parameters.

The requirements are significant. You need Windows 10 or 11 x64, an NVIDIA GPU of the Turing generation or newer with 8 GB of VRAM, driver version 580 or higher, and an SSD. The amount of RAM depends on the model variant: about 32 GB for W4A8, 40 GB for W6A8, and 48 GB for FP8/INT8. For the heavier variants, the authors recommend 64 GB. A laptop without a discrete GPU and with 16 GB of memory will not work.

Who will find it useful

The practical value lies in cost and privacy. New takes don't use up cloud credits: you spend only electricity and your GPU's time. If you work with confidential scripts or with other people's voices under a contract, they stay on your computer. For game studios and course authors, 16 voices in one scene cover most tasks.

The limitations are also clear. The build exists only for Windows and NVIDIA, and its appetite for RAM is high. The studio is distributed through Boosty (card payment) and Telegram (USDT, YooMoney, Telegram Stars), so it can't be called free. The quality of the result depends on Fish Audio Drama 3 and on how clean the sample is, so before a serious project it is worth running a few test takes on your own hardware.

Story

Key Questions Answered

How the studio works?

Python, PyTorch, and ComfyUI are already bundled inside the app, so you don't need to install anything manually. The engine starts automatically or manually, whichever you prefer. The models are downloaded separately from within the program: downloads resume after an interruption, and the files are verified. After that, the studio works without the internet, and your texts and voice samples are not sent anywhere.

What are the key details on Voices, dialogues, and emotions?

Cloning is simple. You upload a recording, select 5–15 seconds of clean speech in it, and use that voice for new text. The voice, together with its transcript, can be saved to a library, so you can return to it in your next video or for one character's lines. The authors remind users to use only recordings they have the rights to. For a scene with several characters, you assign each one its own sample, and up to 16 participants are combined into a single script. You can also start with an ordinary monologue.

What are the key details on Projects and hardware requirements?

A project is saved in a .drama3 file, which holds the script, the settings, and the voice samples themselves. The "Return to studio" button restores the settings of the selected result, and autosave and an archive of drafts let you pick up the work later. For long texts, there is automatic splitting, reading the whole text or line by line, separate takes, and shortening of long pauses. Advanced mode unlocks model selection, seed, and generation parameters. The requirements are significant.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…