NoizAI skills turn a bot into something that sounds like a person

A catalogue of skills, installed one at a time

The repository is a central repository for managing Skills that give a bot a more human voice, and the install model is a single command per skill.

# List skills from GitHub repository
npx skills add NoizAI/skills --list --full-depth

# Install a specific skill from GitHub repository
npx skills add NoizAI/skills --full-depth --skill tts -y

# Install from GitHub repository
npx skills add NoizAI/skills

# Local development (run in this repo directory)
npx skills add . --list --full-depth

The flags do distinct things. Listing with --list and --full-depth shows what is in the repository, adding a single skill with --skill and -y takes the confirmation away so it can run unattended, and the last line is the same thing pointed at the local checkout, which is how you test a skill you are editing.

The top level of the repository is small and consistent with that model. There is a .github directory, a .gitignore, a CONTRIBUTING.md, a Clones.png, a README.md, a Chinese README, an examples directory with an audio subdirectory, and a skills directory. So the skills themselves live in one directory, each with its own SKILL.md, and the repository is a distribution point rather than an application.

The primary language is Python, the last push was on 2026-09-28, and there are no GitHub releases. The licence is not stated, which is worth noting before you depend on it commercially.

Eight skills, and what each one actually does

The skills table lists eight of them with a description, a documentation link and a run command.

tts converts text into speech with Kokoro or the Noiz backend, with a simple mode, timeline-aligned rendering, precise duration control, and reference audio voice cloning. characteristic-voice is the one that makes generated speech feel companion-like, using fillers, emotional tuning and preset speaking styles. sound-fx generates any sound effect from a text description, animals, ambience, cartoon sounds, sci-fi and more, in one command, one to 30 seconds, with WAV, MP3 and FLAC output.

text-to-music generates a full song with vocals from a music description and lyrics, your own or written for you, or covers an existing song in a new style with its lyrics auto-recognized, with two variants per request and MP3 output. music-video builds a 24fps music video from an existing song, either as a p5.js animation or by assembling local stills and clips.

The remaining three are the ones that reach further. video-translation translates and dubs videos from one language to another, replacing the original audio with TTS while keeping the video intact. daily-news-caster fetches the latest real-time news and generates a dual-host conversational podcast with audio. And chat-with-anyone chats with any real person or fictional character in their own voice, by automatically finding their speech online, extracting a clean reference sample, and generating audio replies.

That last one deserves care rather than enthusiasm. The mechanism is built around locating a person's public speech and extracting a reference sample from it, which is a technique whose acceptability depends entirely on whether the person has agreed to it.

Local by default, with a hosted backend behind a key

The highlights list four properties, and the first one sets the tone: secure and local-first, running skills on your own machine to keep sensitive text and assets localized.

That default is real rather than aspirational, because the TTS skill has a local backend. The API key section explains the alternative: for the best experience, which the documentation spells out as faster, emotion control and voice cloning, you get a key from the Noiz developer site and set it with:

bash skills/tts/scripts/tts.sh config --set-api-key YOUR_KEY

The key is persisted to ~/.noiz_api_key and loaded automatically. Alternatively you pass --backend kokoro to use the local Kokoro backend instead.

So the choice is between a hosted service that is faster and can clone a voice, and a local model that needs no key and no network. The persistence detail is the one to be aware of: the key is written to a file in your home directory and read back on every run, which is convenient and means the key is on disk in plain text.

The other two highlights are about output rather than privacy. Production-ready voice covers the range from quick TTS generation to timeline-aligned rendering, and one-command delivery to chat platforms generates speech and sends it as a native voice message to Feishu, Telegram or Discord with zero extra code.

The dependency list, and the hybrid mode that keeps media local

The requirements are stated per skill, and they are worth reading because they determine what you can run where.

The music-video skill needs ffmpeg. Code mode also needs Node.js, Google Chrome and git. And then there is the sentence that matters for anyone handling licensed or personal media: hybrid mode reads only local images and clips, it does not upload them or call an image or video generation API.

That is a meaningful distinction, because the obvious use of a video tool is to send footage somewhere. In hybrid mode the clips stay on the machine, and only the composition is local to the same promise.

The repository also ships a quick verify example, which is the fastest way to hear what the voice presets do:

bash skills/characteristic-voice/scripts/speak.sh \
  --preset comfort -t "Hmm... I'm right here." -o comfort.wav

A preset, a line of text and an output file. That is the whole interface for the character skill, and the English audio demos in the readme give you four sample outputs to listen to: breaking news style, mindful calm style, podcast intro style and startup hype style, which is a reasonable sample of the range the presets cover.

The skills are also usable on their own, outside an agent. The scripts under each skill's scripts directory are ordinary command line programs, which means you can wire them into a script or a cron job without an agent in the loop at all.

Where the repository is going, and what is missing

The project trends section is unusual for a skills repository. There is a GitHub star trend chart, from the star history service, and a Git clone trend chart with a Clones.png image at the top of the repository.

The clone trend is the interesting one, because a skills catalogue is a different kind of artefact from a library. A library is installed once and depended on. A skills catalogue is installed per task, and the interesting metric is how many times somebody fetched the repository to pick up one skill. A clone count that grows while the star count stays flat is a signal of a different usage pattern, and publishing it is a way of saying the repository is a distribution point rather than a project people watch.

What is missing is equally worth noting. There is no licence file, and the licence is recorded as unknown, so there is no clear statement of what you may do with the skill code. The documentation for authoring rules, directory conventions and pull request guidance lives in CONTRIBUTING.md, which is the one pointer a contributor needs and a user does not.

For a user, the practical picture is a set of well-scoped command line tools wrapped in the skill format, published by a single team, with a hosted service behind the good parts and a local model behind everything. The readme is bilingual, with the Chinese version linked at the top, and there is a Discord for discussion.

Editorial conclusion

Adopt these skills if you are building a bot or a content pipeline that has to speak, and you want the TTS, dubbing and audio work packaged as installable skills rather than scripts you maintain yourself. Do not adopt the voice cloning skill for a real person without their consent, since the skill is built to find a person's speech online and reproduce it, and that is a decision about the person rather than about the tool. Before you install, check three things: whether you will use the hosted Noiz backend or the local Kokoro one, since the key is what unlocks the faster path and the emotion control, what the audio pipeline needs on the machine, since ffmpeg, Node.js, Chrome and git are all named as requirements, and where the API key is stored, because it is persisted to a file in your home directory rather than to an environment variable.

Frequently asked questions

Official sources

1. Issues
2. NoizAI/skills on GitHub
3. README