# my-neuro: an AI desktop companion with Live2D avatars and voice cloning

> morettt/my-neuro is a workbench for building a personal AI companion: Live2D avatar, cloned voice, long-term memory and local or API-based LLMs. The README claims sub-second conversation latency on fully local inference, but the setup path is aimed at users willing to run several Python services.

**morettt/my-neuro** — This project lets you create your own AI desktop companion with customizable characters and voice conversations that respond in just 1 second. Features include long-term memory, visual recognition, voice cloning and LLM training. Compatible with various Live2D customizations.

- Repository: https://github.com/morettt/my-neuro
- Stars: 1,380 · Forks: 163
- Language: JavaScript
- License: MIT
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/morettt-my-neuro

## What my-neuro actually solves

Most AI companion projects give you a chat window and a personality prompt. my-neuro tries to cover the whole stack around that prompt: an avatar that moves, a voice that can be trained on your own samples, memory that persists between sessions, and a choice between a local language model and a closed API. The README describes the project as a workbench rather than a product, and that framing is accurate. It packages tools so that you assemble the character yourself, module by module.

The intended user is someone who wants a specific character rather than a generic assistant. The README lists voice customization across male and female voices, Live2D model replacement, image recognition triggered by language intent, and long-term memory that records key information about the user. It also ships a default character called 肥牛 (fake neuro), described in the README as deliberately copied from neuro-sama in appearance with a tsundere, mischievous personality. If you want a ready-made character to talk to, that is the path the project offers. If you want to build your own, the same components are exposed for training and replacement.

The project is inspired by neuro-sama, which is why the name and the default character exist at all. That inspiration is worth stating plainly, because it sets expectations: this is a fan-adjacent project that borrows a persona and then hands you the tools to make a different one.

## The component chain behind a one-second reply

The README claims conversation latency under one second when everything runs locally. That number depends on the whole chain being fast, and the repository layout shows how many stages are involved. Separate batch files start each service: 1.ASR.bat for speech recognition, 2.TTS.bat for speech synthesis, 3.bert.bat for the BERT component, and two MemOS files for the memory system, 4.MEMOS-API.bat and 5.MEMOS-WebUI.bat. RAG.bat handles retrieval.

The data flow implied by that layout is: audio comes in, ASR transcribes it, the language model produces a reply, TTS synthesizes speech, the Live2D layer animates the avatar, and the memory system stores and retrieves facts about the user. The README states that subtitle display and audio can be decoupled, so a character can show Chinese subtitles while the TTS model speaks another language. That is a practical detail for anyone training a voice model on non-Chinese audio.

Voice training defaults to GPT-SoVITS, listed in the acknowledgements as an external project. Memory is built on MemOS, also external. The Minecraft integration comes from mindcraft, and the browser automation tool is microsoft/playwright-mcp. my-neuro's own contribution is the orchestration: the batch files, the WebUI control panel, the Electron installer directory, and the glue that lets these separate projects talk to each other. That is a real piece of work, but it also means the reliability of any single feature is bounded by the upstream project behind it.

## Installing my-neuro and getting a first conversation

The README does not give command-line install steps. It points to a deployment tutorial on the project site and to a packaged archive for beginners, with a warning that the file path must not contain Chinese characters, spaces, brackets or similar symbols. The repository also contains an installer.py, an electron-installer directory, a 预览安装器UI.bat preview script, and a 一键更新live-2d.bat updater, which suggests the intended path is running the installer rather than pip-installing by hand.

If you install from source, requirements.txt is the dependency list. Note that torch and torchaudio are commented out with a note that you must install the CUDA version yourself:

```bash
pip install -r requirements.txt
# torch and torchaudio are commented out in requirements.txt;
# install the CUDA build manually before running the ASR and TTS services
```

The services are started by the numbered batch files in the repository root. A minimal order for a first run is ASR, then TTS, then the memory API, since the conversation loop reads from all three:

```bash
1.ASR.bat
2.TTS.bat
4.MEMOS-API.bat
```

After those are running, the WebUI control panel is the entry point for configuration:

```bash
启动 WebUI 控制面板.bat
```

What you should see is the control panel opening in a browser and the individual services reporting that they are listening. The README does not document the ports these services bind to, so check the console output of each batch file rather than assuming a default.

For a fully local setup with no third-party API, the README directs you to the LLM-studio folder, which contains instructions for local model inference and fine-tuning. That folder is the difference between running the packaged character and training your own.

## Where my-neuro breaks down or is the wrong choice

The plan list in the README is the clearest statement of what is unfinished. Real emotion, described as simulating a person's changing emotional state, is unchecked. So are overseas livestream platform integration, color-changing based on mood, and free movement of the avatar around the screen. The README says the project has implemented close to 30% of its intended functionality and that the human-like emotional layer is the next major piece. Anyone adopting my-neuro for the emotional continuity it promises is adopting a roadmap, not a finished feature.

The latency claim is also conditional. It applies to fully local inference, and it depends on your hardware hosting ASR, TTS and the language model simultaneously. A machine that runs a small local model comfortably may not run it alongside GPT-SoVITS and FunASR without contention. The README gives no hardware requirements, which is a gap for a project whose headline number is a latency figure.

The dependency surface is the third constraint. requirements.txt pulls in fastapi, funasr, sentence-transformers, qdrant-client, neo4j, rank-bm25, transformers, PyQt5 and pyinstaller, among others. Several of these are large. If you only want a text chatbot with a personality prompt, my-neuro is the wrong tool: you would be installing a speech pipeline, a vector database and a graph database to get a chat loop. The project is for people who want the avatar, the voice and the memory together, and are willing to maintain the stack that makes them work.

## my-neuro versus a plain LLM chat front end

The obvious alternative is a general chat interface pointed at a local model, with a system prompt for the character. That approach gives you text conversation and nothing else: no avatar, no synthesized voice, no persistent memory store, no visual recognition. It also installs in minutes and has one moving part.

The difference in approach matters more than the feature count. A chat front end treats personality as prompt text. my-neuro treats it as a set of trained artifacts: a voice model trained through GPT-SoVITS, a memory store managed by MemOS, a Live2D model you can swap, and optional fine-tuning through the LLM-studio folder. The README's own framing is that the personality, appearance, voice and emotional variation are all decided by you, with the project supplying the modules.

That is a heavier commitment in both setup and ongoing maintenance. It is also the only way to get a character that sounds like a specific voice and remembers you across sessions without you pasting context every time. If the voice and the memory are not the point for you, the simpler front end wins on every axis except appearance.

## Maintenance, licensing and what upgrading costs

The repository is not archived, and the last push was on 2026-09-09. Releases are frequent: v6.7.2 on 2026-09-09, v6.7.1 on 2026-09-07, and an installer package labeled exe3.0 on 2026-09-07. Frequent releases at this cadence mean the upgrade path is real but also that configuration files and batch scripts may change between versions. The repository includes 一键更新live-2d.bat and update.py, so the project provides its own update mechanism rather than relying on you to pull and rebuild.

my-neuro itself is MIT licensed, which permits commercial and private use with the usual attribution requirement. The components it orchestrates are separate projects with their own licenses, and the README lists them: GPT-SoVITS for TTS, MemOS for memory, mindcraft for Minecraft, and playwright-mcp for browser automation. Your obligations depend on those upstream licenses, not on my-neuro's, and the README does not summarize them. If you plan to distribute anything built on this stack, check each component individually.

The maintenance cost is the number of services. Five batch files start five processes, and each has its own dependencies and failure modes. Expect to debug the memory API separately from the TTS service, and expect a torch or CUDA mismatch to surface as a failed ASR start rather than a clear error message.

## Conclusion

my-neuro is for people who want to assemble a personal AI character and are comfortable running Python services, batch files and a local TTS pipeline. It is not for anyone who wants a single installer that behaves like a finished desktop app, and the README itself states the emotional-state system is still unchecked on the plan list. Before adopting it, verify that your GPU can host the ASR, TTS and language model components at once, and read the deployment tutorial on the project site, since the repository ships batch files rather than a documented CLI.

## FAQ

### How do I install my-neuro?

The README points to a deployment tutorial on the project site and offers a packaged archive for beginners, with the warning that the file path must not contain Chinese characters, spaces or brackets. Installing from source means running pip install -r requirements.txt and installing the CUDA build of torch manually, since those lines are commented out.

### Does my-neuro run fully locally?

Yes, the README states that fully local inference is supported and that the LLM-studio folder contains guidance for local model inference and fine-tuning. It also supports closed model APIs as an alternative. The sub-one-second latency claim applies to the fully local path.

### Which TTS project does my-neuro use for voice cloning?

The README lists GPT-SoVITS as the default TTS training project, and it appears in the acknowledgements as an external open source project. Voice customization across male and female voices and different character timbres is listed as a checked feature.

### Is the emotional state feature of my-neuro finished?

No. The plan list marks real emotion, described as simulating a person's changing emotional state, as unchecked, and the README says the project has implemented close to 30% of its intended functionality with the emotional layer as the next major piece.

## Sources

- [Issues](https://github.com/morettt/my-neuro/issues)
- [License: MIT](https://github.com/morettt/my-neuro/blob/main/LICENSE)
- [morettt/my-neuro on GitHub](https://github.com/morettt/my-neuro)
- [README](https://github.com/morettt/my-neuro/blob/main/README.md)
- [Releases](https://github.com/morettt/my-neuro/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/morettt-my-neuro