MOSS-TTS

MOSS-TTS

Reads text aloud in eight languages with zero-shot voice cloning, long-form speech and control over pronunciation and duration. For audio producers and developers who build voice features.

Made by OpenMOSS-Team on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MOSS-TTS is the flagship text to speech model of the MOSS-TTS Family from MOSI.AI and the OpenMOSS team. It focuses on natural sounding speech, voice cloning from a short reference, and long stretches of audio. It also lets you steer pinyin, phonemes and duration, and it handles text that mixes languages. Its record lists 8.5B parameters.

Who it suits: It suits people who turn text into spoken audio, from audio producers to developers adding a voice to a product. The wider family adds models for dialogue, voice design, real-time replies and sound effects.

Industry | Who uses it | How they use it | Example
Media | Podcast and audiobook producers | Generate long form narration from a script | Reading a long article aloud in one chosen voice
Games and film | Sound designers | Create scene sounds with the sound effect model in the family | A city street scene of a set length
Voice products | Developers of voice agents | Pair the real-time model with a text model for fast replies | A spoken answer that begins after about 180 ms
Localization | Teams with several languages | Synthesize text across the listed languages | A sentence that switches between Chinese and English

You can skip it if you need speech recognition or text-only work, since it only produces audio.

Where to use it:

• Python on Windows, macOS or Linux, after downloading the files
• A voice agent that pairs a text model with the real-time model
• Audio production for podcasts, dialogue, games and film

Use case | Who | What they get back
Narrate a long text | Producers | Stable speech over many minutes
Clone a voice you may use | Creators | New speech in a voice taken from a short reference
Make a spoken dialogue | Content teams | Multi-speaker conversation audio

How to set it up:

• Python
• Space for the weights, about 17 GB in BF16 or about 4.2 GB at 4-bit by estimate
• A voice you have the right to use, if you plan to clone one

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Open the model card and follow its steps to run the model.

pip install -U huggingface_hub
hf download OpenMOSS-Team/MOSS-TTS

How to use it: You give the model text, and optionally a reference voice, and it returns speech audio. The model card holds the run steps. The family has five models that work alone or together in one pipeline.

Tool | What it does
MOSS-TTS | The flagship model for voice cloning, long speech and multilingual synthesis
MOSS-TTSD | Generates spoken dialogue with several speakers
MOSS-VoiceGenerator | Designs a voice from a text prompt, with no reference speech
MOSS-TTS-Realtime | Answers in speech with low latency for voice agents
MOSS-SoundEffect | Generates sound effects with a chosen duration

Example: Give MOSS-TTS a paragraph of text and a short clip of a voice you have the right to use. It speaks the paragraph in that voice, and you can adjust pronunciation and timing.

For newcomers and experts: This is software that reads written text out loud in a natural voice. For a first try, download the files with the two commands above, then follow the model card with one short sentence. Use only voices you have the right to use.

Underneath, the family trains two architectures under one setup. MossTTSDelay favors long-context stability and inference speed, and MossTTSLocal favors lightweight flexibility for streaming systems. The real-time model is a separate design for voice agents that uses earlier text and the user's audio across turns.

Text and an optional reference voice
MOSS-TTS model generates the speech
Pinyin, phoneme and duration controls apply
Speech audio comes back

Good to know: The family covers Chinese, English, German, Spanish, French, Japanese, Italian and Hebrew. The real-time model reports its first audio byte after 180 ms.

Title: MOSS-TTS-Realtime latency in milliseconds
Time to first byte: 180 ms
First text sentence plus first byte: 377 ms

• License: Apache 2.0
• Cost to run: none named for running the files yourself, and the card also points to a hosted studio and API without giving prices
• Data it reaches: downloads model files from Hugging Face, then works on the text and reference audio you give it
• Not for: cloning a voice you have no right to use

On Hugging Face, 13 November 2026

• Task: text to speech
• Parameters: 8.5B
• Weights: about 17 GB in BF16; about 4.2 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,748
• Likes: 438
• License: apache-2.0
• Languages: zh, en, de, es, fr, ja, it, he
• Quantized versions: 1
• Fine-tunes: 1
• Tags: safetensors, moss_tts_delay, text-to-speech

Links

• Model: https://huggingface.co/OpenMOSS-Team/MOSS-TTS
• Files: https://huggingface.co/OpenMOSS-Team/MOSS-TTS/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download OpenMOSS-Team/MOSS-TTS

Then follow the usage in its model card. Nothing in this kit runs by itself.