MOSS-TTSD

OpenMOSS's 8B model for spoken dialogue, with up to 5 speakers and 60 minutes in one pass, plus voice cloning. Apache 2.0.

In short. MOSS-TTSD turns a script with several speakers into a spoken conversation, such as a podcast, an interview or a sports commentary, in one go. You give it a short recording of each speaker and it keeps each voice consistent for up to an hour. It is free to use in a business.

Summary

MOSS-TTSD v1.0 is the dialogue specialist of the MOSS-TTS family. It is an 8B model on the same MossTTSDelay design as MOSS-TTS, and it turns a script with speaker tags such as [S1] and [S2] into a spoken conversation. According to the card it handles 1 to 5 speakers, up to 60 minutes of audio in a single pass, and zero-shot voice cloning from references of 3 to 10 seconds. It is tuned for podcasts, interviews, audiobooks with a narrator and characters, sports and esports commentary, dubbing and crosstalk. The family card lists 20 languages and the TTSD section describes cross-lingual performance as robust in major languages including Chinese, English, Japanese and European languages. You supply a reference clip and transcript for each speaker, and the model continues in each voice.

OpenMOSS's own tests on Chinese and English dialogue list MOSS-TTSD v1.0 with the best speaker similarity and speaker-attribution accuracy among the open models, and the lowest Chinese word error rate, 4.85%. VibeVoice 7B has a slightly lower English word error rate, 9.46% against 9.88%. Against proprietary systems (Eleven v3, Gemini 2.5 Pro TTS and Doubao Podcast, with MOSS-TTSD tested on voices from each system), the card lists higher speaker similarity and attribution accuracy for MOSS-TTSD but a higher word error rate in every comparison. The README adds that annotators compared pairs of samples and preferred MOSS-TTSD to Doubao and Gemini 2.5 Pro, but it shows those results only as charts. All of this is the developer's own testing.

The weights are Apache 2.0, 16.7 GB in BF16, and load with trust_remote_code on Transformers 5. vLLM-Omni lists support for MOSS-TTSD. The technical report is arXiv 2603.19739. In the 30 days to 11 November 2026, Hugging Face counted about 11,800 downloads, far fewer than MOSS-TTS itself.

Good for

  • Podcast-style audio and interviews from a written script with two to five speakers.
  • Long audio of up to an hour, such as audiobooks with a narrator and characters, or commentary.
  • Dubbing and crosstalk, with each voice cloned from a 3 to 10 second reference.
  • A commercial dialogue pipeline under Apache 2.0, in place of research-only options such as VibeVoice.

Watch out for

  • All scores are OpenMOSS's own, and the human-preference results are shown only as charts.
  • Word error rates are higher than the proprietary systems it is compared with, even though similarity and attribution scores are better.
  • Language coverage beyond Chinese and English is described only in general terms, so test the language you need.
  • It continues from a reference clip and transcript for every speaker, so you need clean samples and scripts for them, and an 8B model needs a large GPU.

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of MOSS-TTSD. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
v1.0 (8B) BF16 8.36B 17 GB 19 GB Gaming GPU

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is MOSS-TTSD?
MOSS-TTSD turns a script with several speakers into a spoken conversation, such as a podcast, an interview or a sports commentary, in one go. You give it a short recording of each speaker and it keeps each voice consistent for up to an hour. It is free to use in a business.
Can I use MOSS-TTSD commercially?
Yes. MOSS-TTSD is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is MOSS-TTSD open source?
MOSS-TTSD is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run MOSS-TTSD on my own computer?
Yes, on a good graphics card. The smallest variant, v1.0 (8B), needs about 19 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was MOSS-TTSD released?
MOSS-TTSD was released by OpenMOSS on February 10, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
MOSS-TTS OpenMOSS 4.55B – 8.49B – Apache 2.0 (Permissive) 10 GB Laptop or phone
VibeVoice-TTS Microsoft 1.02B – 2.7B 64K MIT (Permissive) 2.35 GB Laptop or phone
Drama 3 Fish Audio 35.9B 3B active – Fish Audio Community 1.0 (Conditional) 25 GB Gaming GPU
Higgs TTS 3 Boson AI 4.65B – Boson Higgs Non-Commercial (Non-commercial) 11 GB Laptop or phone