Ming-3.0-tts-0.5B by inclusionAI

Ming-3.0-tts-0.5B by inclusionAI

Turns text into speech in seven languages, including Chinese and English. It can copy a voice from a short clip, design one from a description, and control emotion, speed, pitch and volume.

Made by inclusionAI 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: Ming-3.0-tts-0.5B is a text to speech checkpoint from inclusionAI. It reads written text aloud in Chinese, English, German, Spanish, French, Korean and Japanese, plus Sichuanese and Cantonese dialects.

It can copy a voice from a short reference recording or build a voice from a written description of the speaker. The model sits on a Qwen2.5-0.5B backbone.

Who it suits: Developers and audio makers who want a speech generator they can run themselves and steer with tags, emotion and speed settings.

Industry | Who uses it | How they use it | Example
Software | Developers adding voice output | Call the Python API from an app | Read assistant replies aloud in Chinese or English
Media | Video and audio editors | Generate narration in a chosen delivery style | Voice-over with breath or laughter tags
Education | Language course makers | Produce spoken lines in several languages | Practice sentences in Korean or French
Accessibility | Teams building read-aloud features | Turn articles into speech with a built-in voice | Narrated reading of a long text

Skip it if you want a ready-made app, because this repository holds only the checkpoint and running it needs separate code.

Where to use it:

• A terminal, through the infer.py script
• A Python program, through the MingTTS class
• Windows, macOS or Linux with Python, to download the files
• NVIDIA GPU machines, where the authors say they tested it with CUDA 12.4

Use case | Who | What they get back
Voice cloning from a short clip | Creators with permission to use that voice | A wav file in the reference voice
Voice design from a description | Game and app developers | A new speaker built from your wording
Built-in voices | Anyone who needs quick narration | Speech in UK_George, US_Emma, 灵小柔 or 灵楷

Only clone a voice you have the right to use.

How to set it up:

• Python and pip
• The separate Ming-3.0-tts code from GitHub, which holds the inference scripts and requirements
• A GPU is not named as required, but the authors tested on NVIDIA H800 and H20 cards

1. Install the Hugging Face download tools and fetch the checkpoint.
2. Clone the Ming-3.0-tts code from GitHub and open its folder.
3. Install its requirements.
4. Optionally install pynini for number and date normalization, or pass a flag to skip it.

pip install -U huggingface_hub
hf download inclusionAI/Ming-3.0-tts-0.5B

pip install -r requirements.txt

conda install -c conda-forge pynini==2.1.6

How to use it: Run the infer.py script with a task, a language, your text and, for cloning, a reference recording with its transcript. The output option sets where the wav file goes.

python infer.py \
  --model_path inclusionAI/Ming-3.0-tts-0.5B \
  --task TTS \
  --language_id Chinese \
  --text "傍晚的风从山口吹过来，远处的云慢慢染成了金色。" \
  --prompt_wav_path data/wavs/10002287-00000094.wav \
  --prompt_text "在此奉劝大家别乱打美白针。" \
  --use_spk_emb \
  --output output/tts.wav

Tool | What it does
infer.py | Command line script that runs a task and writes audio
ming_tts.MingTTS | Python class with a synthesize call for the same tasks
TTS | Speaks text, optionally in a reference voice
VOICE_DESIGN | Builds a voice from a written speaker description
FREEFORM | Follows a free-form instruction about delivery and style
MULTILINGUAL | Speaks German, Spanish, French, Korean or Japanese, which need a language ID
EMOTION and DIALECT | Control emotion, or Sichuanese and Cantonese speech

Example: You give the command above a Chinese sentence about evening wind over a mountain pass, plus a short recording and its transcript. The model speaks the sentence in the voice from the recording and saves it as a wav file in the output folder.

For newcomers and experts: If you are new to this, treat it as a voice generator. Download the files first, then follow the GitHub instructions to turn one short English sentence into a wav file and listen to it.

If you know the field, the checkpoint bundles a Qwen2.5-0.5B based backbone, a DiT head, an audio tokenizer, a speaker encoder and a text tokenizer. Output is 44.1 kHz. Text normalization uses FST grammars through pynini, and flash-attn is optional because inference falls back to PyTorch SDPA.

Pick a task and a language
Add your text and an optional reference clip
Run infer.py or the MingTTS class
Receive a wav file

Good to know:

• Weights: about 3 GB in BF16, according to Hugging Face
• Freeform runs as a separate task, so it will not mix with the voice design, built-in voice, emotion or dialect options
• Fourteen inline tags cover breaths, coughs, sighs, laughter, hesitation and surprise sounds
• License: MIT
• Cost to run: none named
• Data it reaches: the checkpoint it downloads, your text and any reference recording you supply, and the wav file it writes
• Not for: standalone use from this repository, since the inference code lives elsewhere

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: transformers
• Parameters: 1.5B
• Weights: about 3 GB in BF16; about 740 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 14,756
• Likes: 225
• License: mit
• Languages: zh, en, de, es, fr, ko, ja
• Quantized versions: 17
• Fine-tunes: 4
• Tags: transformers, onnx, safetensors, dense, text-generation, audio, text-to-speech, voice-cloning, multilingual, speech-generation

Links

• Model: https://huggingface.co/inclusionAI/Ming-3.0-tts-0.5B
• Files: https://huggingface.co/inclusionAI/Ming-3.0-tts-0.5B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download inclusionAI/Ming-3.0-tts-0.5B

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