RepoPicks Radar

Updated every day around 23:30 UTC

Guides

MOSS-TTSD

OpenMOSS/MOSS-TTSDReviewed by AI

Checked by RepoPicks Radar on

What is MOSS-TTSD?

MOSS-TTSD turns a written dialogue into spoken conversation: 1 to 5 speakers, up to 60 minutes in one session, in 20 languages, with each voice copied from a short reference recording. The authors aim it at podcasts, audiobooks, dubbing and sports or esports commentary. The model has 8 billion parameters and the documented setup expects Python 3.12, an NVIDIA graphics card and commands in a terminal; the authors also host a free online demo you can try first. Free under Apache-2.0, with the model files downloaded from Hugging Face.

MOSS-TTSD is free and open source (Apache-2.0) and can replace ElevenLabs, NotebookLM, and Descript.

How easy is it to install MOSS-TTSD?

A · download and open · Web

  • Windows D
  • Linux D
  • Web A

The easiest way in is the web at grade A: the authors keep a free Hugging Face Space, so you open a page and listen. Running it yourself is D on Linux and Windows: `conda create -n moss_ttsd python=3.12`, `pip install -r requirements.txt` and `pip install flash-attn`, where requirements.txt pins the CUDA builds `torch==2.9.1+cu128` and `torchaudio==2.9.1+cu128`, so an NVIDIA card is assumed; then you write a Python script or a JSONL file and run `python inference.py`. macOS is not graded because the pinned CUDA packages and flash-attn are not for Apple machines.

MOSS-TTSD on GitHub

Stars
1,420on Nov 10, 2026
Downloads
—all releases
Growth
+4stars in 6 days

Stars over time

1,4171,4181,4191,4201,421 Nov 7Nov 8Nov 9Nov 10 1,4171,4181,4191,4201,421 Nov 7Nov 9

Facts

Runs on
Windows, Linux, Web
You need to install
Python, NVIDIA graphics card
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 the hardware: an 8B model needs a graphics card with enough memory, and the README warns about "VRAM fragmentation" if "GPU memory is tight".
License
Apache-2.0 (Permissive)
Last change on GitHub
Sep 6, 2026

MOSS-TTSD risk signals (8)

  • AttentionDependencies with known vulnerabilities

    Pinned dependencies with known vulnerabilities (OSV): torch: GHSA-63cw-57p8-fm3p, GHSA-qfhq-4f3w-5fph, GHSA-rrmf-rvhw-rf47; transformers: GHSA-29pf-2h5f-8g72, GHSA-fgcw-684q-jj6r, GHSA-x9r9-c232-4q39; orjson: GHSA-hx9q-6w63-j58v, PYSEC-2026-107; gradio: GHSA-39mp-8hj3-5c49, GHSA-6655-8ph2-63j3, GHSA-7hp7-4p35-3cx2

    automatic check

  • Attentioncopies voices from short samples

    The project description promises "zero-shot voice cloning" and the recommended mode is `--mode voice_clone_and_continuation`, fed with `prompt_audio_speaker1` to `prompt_audio_speaker5`. The README's own Usage Disclaimer says: "Users must not use this model for unauthorized voice cloning, impersonation, fraud, scams, deepfakes, or any illegal activities" — the wish is written down, but nothing in the code checks consent or marks the audio it produces.

    AI review

  • Attentionreadme example invents a talk between real people

    The Quick Start example clones two voices from `asset/reference_02_s1.wav` and `asset/reference_02_s2.wav` whose transcripts are "In short, we embarked on a mission to make America great again for all Americans" and "NVIDIA reinvented computing for the first time after 60 years", then makes them hold an invented argument about China that ends "It's a disaster, Jensen. A total disaster." The project's first demonstration is words put into the mouths of recognisable public figures.

    AI review

  • Attentionruns code downloaded with the model

    The documented calls are `AutoProcessor.from_pretrained(..., trust_remote_code=True)` and `AutoModel.from_pretrained(..., trust_remote_code=True)`, and the server is started with `sglang serve ... --trust-remote-code`. That setting lets Python files stored next to the weights on Hugging Face run on your machine.

    AI review

  • Attentionthe fast server listens on every network address

    The SGLang instructions are `sglang serve --model-path <path-to-fused-model> --delay-pattern --trust-remote-code --port 30000 --host 0.0.0.0`. Host 0.0.0.0 means any machine that can reach yours may send requests to that port, and no password or token is mentioned.

    AI review

  • Infospeed setup uses a forked engine

    To get the faster path the README tells you to `git clone https://github.com/OpenMOSS/sglang -b moss-ttsd-v1.0-with-cat` and `pip install ./sglang/python[all]` — the project's own branch of SGLang, which it calls "the deeply extended SGLang from OpenMOSS", not the official release.

    AI review

  • Infothe comparison tables are the authors own

    The scores against Eleven V3, gemini-2.5-pro-preview-tts, Doubao_Podcast and Vibevoice come from the authors' own TTSD-eval framework and their own annotators ("we compute Elo ratings"), published alongside the paper arXiv:2603.19739. No independent test is cited.

    AI review

  • Infono versions published on github

    There are zero GitHub releases; versions (v0.5, v0.7, v1.0) exist only as Hugging Face model repositories and News entries, with the older code kept in a `legacy/` folder. The last commit is 2026-09-06 (63 days before this check) and there are 53 open issues against 1418 stars.

    AI review

Evidence is shown in its original language.

What Radar could not check about MOSS-TTSD

  • Real disk and graphics-memory use, since the model is 8 billion parameters and the estimate in the checks is a heuristic, not a measurement
  • Whether 60-minute audio really stays stable, as that figure is the authors' own claim
  • How good the result sounds in the 18 languages beyond Chinese and English, which the published tables do not cover
  • What the authors' Hugging Face Space does with the reference audio and the script you upload
  • Which Python files come down with the weights and run on your machine through trust_remote_code

Similar projects

The RepoPicks app

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

Before you download

See the release page on GitHub