RepoPicks Radar

Updated every day around 23:30 UTC

Guides

MOSS-TTS-Nano

OpenMOSS/MOSS-TTS-NanoReviewed by AI

Checked by RepoPicks Radar on

What is MOSS-TTS-Nano?

MOSS-TTS-Nano is a small text-to-speech model (0.1B parameters) you run yourself. Give it a few seconds of someone speaking and it reads your text in that voice, in 20 languages, as 48 kHz stereo audio. The authors say it is fast enough for real time on a 4-core processor with no graphics card, and that the ONNX version runs on a single core of a MacBook Air M4. You install Python packages and open a page on your own machine, or try the authors' free online demo first. The model files download on first run.

MOSS-TTS-Nano is free and open source (Apache-2.0) and can replace ElevenLabs, Murf, and Fish Audio.

How easy is it to install MOSS-TTS-Nano?

A · download and open · Web

  • Windows C
  • macOS C
  • Linux C
  • Web A

The easiest way in is the web at grade A: the authors keep a free online demo and a Hugging Face Space, so you open a page and type text. Installing it is C on Windows, macOS and Linux: Python 3.10 or newer, a Conda environment, `pip install -r requirements.txt` and `pip install -e .`, then `python app_onnx.py` and open http://127.0.0.1:18083 — no graphics card needed. The README itself warns that one dependency often fails ("If `WeTextProcessing` or `pynini` fails to install"), which can turn those few commands into a longer fight.

MOSS-TTS-Nano on GitHub

Stars
4,657on Nov 10, 2026
Downloads
—all releases
Growth
+44stars in 6 days

Stars over time

4,6304,6404,6504,660 Nov 6Nov 7Nov 8Nov 9Nov 10 4,6304,6404,6504,660 Nov 6Nov 8Nov 10

Facts

Runs on
Windows, macOS, Linux, Web
You need to install
Python
To build from source
Python, Graphics card
Disk space
950 MB–4.3 GB · estimate
Accounts or keys
None
Cost
Free · Free software under Apache-2.0 and free model files. The cost is your own machine plus roughly 1 to 4.5 GB of disk for the downloaded models.
License
Apache-2.0 (Permissive)
Last change on GitHub
Sep 6, 2026

MOSS-TTS-Nano risk signals (7)

  • AttentionDependencies with known vulnerabilities

    Pinned dependencies with known vulnerabilities (OSV): torch: GHSA-3749-ghw9-m3mg, GHSA-63cw-57p8-fm3p, GHSA-887c-mr87-cxwp; transformers: GHSA-29pf-2h5f-8g72, GHSA-69w3-r845-3855, GHSA-fgcw-684q-jj6r

    automatic check

  • Attentioncopies a voice from a short clip

    The README calls voice cloning "the main recommended workflow for MOSS-TTS-Nano" and the example is `python infer.py --prompt-audio-path assets/audio/zh_1.wav --text "..."`, so any recording you have becomes a voice the model can speak with. Nothing in this package mentions consent rules, a disclaimer or a mark on the generated audio.

    AI review

  • Attentioninstall step pulls code from a personal fork

    The recommended Conda setup runs `pip install git+https://github.com/WhizZest/WeTextProcessing.git`, a copy of the text-normalization package held in an individual's account rather than the original project or PyPI, and the README also points to "a community-tested example, see Issue #6" for finding a `pynini` wheel.

    AI review

  • Inforeadme license section contradicts the license file

    README License section: "This repository will follow the license specified in the root `LICENSE` file. If you are reading this before that file is published, please treat the repository as **not yet licensed for redistribution**." The root LICENSE file is in fact the Apache License 2.0 and GitHub reports Apache-2.0, so this looks like leftover text from the first release.

    AI review

  • Infodownloads models on first run

    README: "If `--model-dir` is omitted, the script automatically checks `./models`. When the model files are missing, it downloads them on first run from" the Hugging Face repositories MOSS-TTS-Nano-100M-ONNX and MOSS-Audio-Tokenizer-Nano-ONNX. The checks estimate 950 MB to 4450 MB of disk.

    AI review

  • Infospeed and quality numbers come from the authors

    The "nearly 2x" efficiency, the single-core MacBook Air M4 result and the "best overall reconstruction quality" tables are all the authors' own measurements from their README and papers (arXiv 2603.18090 and 2602.10934). The README also notes the model is 0.1B parameters, far smaller than the 8B flagship MOSS-TTS in the same family.

    AI review

  • Infoyoung project at version 0 1 0

    Created 2026-04-10, pyproject.toml says version "0.1.0", there are no GitHub releases at all, and the last commit is from 2026-09-06 (63 days before this check), with 28 open issues against 4645 stars.

    AI review

Evidence is shown in its original language.

What Radar could not check about MOSS-TTS-Nano

  • How natural the cloned voice sounds for your language, since the authors' own tests focus on Chinese and English
  • Whether it really keeps up in real time on your processor, as the CPU claims are the authors' measurements on their machines
  • Real disk use after the model files are downloaded, estimated at 950 MB to 4450 MB but not stated by the project
  • What the authors' online demo and Hugging Face Space do with the reference audio and text you upload
  • Whether the files fetched from Hugging Face on first run match the model the paper describes, as no checksum is given

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub