MOSS-TTS-v1.5

MOSS-TTS-v1.5

Reads text aloud in 31 languages and clones a voice from a short reference, with language tags and explicit pause markers. For narrators, dubbing teams and developers who add speech to products.

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-v1.5 is a text to speech model with 8.5B parameters, continued from MOSS-TTS 1.0 by the OpenMOSS team. It keeps what 1.0 could do: cloning a voice from a short reference without training, long speech, duration control at token level, pinyin or IPA pronunciation control, and code-switching. Version 1.5 adds language tags, steadier cloning and explicit pause markers.

Who it suits: It suits people who turn text into spoken audio in many languages and want fine control over how it sounds. That includes narrators, dubbing teams and developers adding speech to a product.

Industry | Who uses it | How they use it | Example
Media | Narration and dubbing teams | Clone a voice from a reference and generate long speech | A narration with a set pause before a poem title
Localization | Multilingual product teams | Tag the language of each line so pronunciation stays right | Passing French as the language for a French sentence
Education | Makers of language lessons | Control pronunciation with pinyin or IPA | A Chinese lesson clip with chosen pronunciations
Voice products | Developers | Add synthesized speech to an app | A script that builds a message and saves the generated audio

You can skip it if you need speech recognition or text-only work, or if your machine cannot hold an 8.5B model.

Where to use it:

• Python on Windows, macOS or Linux, in a clean environment
• A GPU machine, with FlashAttention 2 as an optional speed-up when the hardware supports it
• Audio production pipelines for narration, dubbing and voice features

Use case | Who | What they get back
Narrate in a cloned voice | Narrators | Speech that sounds like the reference voice
Speak a tagged language | Localization teams | Pronunciation fitted to the language you name
Pause where you choose | Audio editors | A pause of the length you write into the text

How to set it up:

• Conda or another way to make a clean Python 3.12 environment
• Transformers 5.0.0, which the model card recommends to avoid conflicts
• 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. Create and activate a clean environment.
2. Clone the MOSS-TTS project from GitHub and install it with pip in editable mode, as the model card's Quick Start shows.
3. Optionally add the FlashAttention 2 extra for speed and lower GPU memory.
4. Download the model files.

conda create -n moss-tts python=3.12 -y
conda activate moss-tts

To fetch the model files yourself:

pip install -U huggingface_hub
hf download OpenMOSS-Team/MOSS-TTS-v1.5

How to use it: Build the user message with the processor and set the language when you know it. With the language named, 1.5 beats 1.0 on almost all supported languages.

processor.build_user_message(text=text_fr, language="French")

Inline pause markers go straight into the text, as in this line.

我今天学习了一首中国的古诗，它的名字是[pause 3.2s]静夜思！

Tool | What it does
processor.build_user_message | Builds the message that carries your text and options
language | Tags the language of the text to improve synthesis
[pause 3.2s] | Inserts an explicit pause of 3.2 seconds at that spot
hf download | Fetches the model files to your computer

Example: Write a Chinese sentence about a classic poem and place a pause marker of 3.2 seconds before the poem's name. The generated speech holds that silence, then says the name.

For newcomers and experts: It reads text out loud and can imitate a voice from a short sample. For a first try, set up the environment from the steps above, download the files, and speak one short sentence with the language set. Use only voices you have the right to use.

Underneath, 1.5 continues training from 1.0 with extra multilingual data, and the card lists cloning that varies less between runs. It also follows punctuation pauses more closely in long sentences and handles a long reference with short text better. Leaving the language empty can help some languages and slightly hurt others compared with 1.0.

Text, language tag and optional reference voice
The processor builds the user message
MOSS-TTS-v1.5 generates the speech
Pause markers and punctuation shape the timing
Speech audio comes back

Good to know: The model supports 31 languages, up from 20 in version 1.0. New ones include Thai, Swahili and Cantonese.

Title: Supported languages by version
MOSS-TTS 1.0: 20 languages
MOSS-TTS-v1.5: 31 languages

• 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 and the project code from GitHub, 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: 80,361
• Likes: 220
• License: apache-2.0
• Languages: zh, yue, en, ar, cs, da, de, nl
• Quantized versions: 8
• Fine-tunes: 5
• Tags: safetensors, moss_tts_delay, text-to-speech

Links

• Model: https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5
• Files: https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5/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-v1.5

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