--- title: Ming-3.0-tts-0.5B canonical_url: "https://www.modelscope.ai/models/inclusionAI/Ming-3.0-tts-0.5B" md_url: "https://www.modelscope.ai/models/inclusionAI/Ming-3.0-tts-0.5B.md" repository: inclusionAI/Ming-3.0-tts-0.5B last_updated: 2026-11-03 license: mit pipeline_tag: text-to-speech tasks: - text-to-speech model_type: - dense architectures: - BailingMMNativeForConditionalGeneration parameters: 1.5B tensor_type: - BF16 - F32 library_name: - onnx - safetensors - pytorch frameworks: - pytorch language: - zh - en - de - es - fr - ko - ja downloads: 2119 stars: 0 tags: - audio - text-to-speech - voice-cloning - multilingual - speech-generation --- # Ming-3.0-tts-0.5B > Ming-3.0-tts-0.5B - An open-source model by inclusionAI on ModelScope. Ming-3.0-tts-0.5B Dense 0.5B multilingual speech synthesis with paralinguistic and free-form control. inclusionAI/Ming-3.0-tts-0.5B is a 1.5B-parameter text-to-speech model on ModelScope. licensed under mit. - **Repository**: inclusionAI/Ming-3.0-tts-0.5B - **License**: mit - **Tasks**: text-to-speech - **Parameters**: 1.5B - **Tags**: audio, text-to-speech, voice-cloning, multilingual, speech-generation - **Downloads**: 2119 - **Stars**: 0 - **Last updated**: 2026-11-03 Source: https://www.modelscope.ai/models/inclusionAI/Ming-3.0-tts-0.5B ---
Ming-3.0-tts Logo

Ming-3.0-tts-0.5B

Dense 0.5B multilingual speech synthesis with paralinguistic and free-form control.

GitHub | Blog & Demos | ModelScope

This repository contains **Ming-3.0-tts-0.5B**, the dense Ming-3.0-tts checkpoint built on Qwen2.5-0.5B. The checkpoint includes the backbone, DiT head, audio tokenizer, speaker encoder and text tokenizer. > **Inference requires the [Ming-3.0-tts GitHub repository](https://github.com/inclusionAI/Ming-3.0-tts).** This Hugging Face repository is for checkpoint distribution and does not provide a standalone inference setup. Clone the GitHub repository and follow its README for the current dependencies, inference code and usage instructions. The CLI scripts and Python APIs referenced below are provided by the GitHub repository. ## Model Details | Item | Value | | --- | --- | | Architecture | Dense, Qwen2.5-0.5B backbone | | Parameters | 0.5B | | Languages | Chinese, English, German, Spanish, French, Korean, Japanese | | Dialects | Sichuanese, Cantonese | | Python API | `ming_tts.MingTTS` | | CLI | `infer.py` | | Sample rate | 44.1 kHz | ## Capabilities - Zero-shot voice cloning from a short reference clip - Voice Design from a natural-language speaker description - Free-form control of delivery and speaking style - Built-in Chinese and English voices - 14 inline tags for breath, cough, sigh, laughter, hesitation, surprise and question/confirmation sounds in Chinese or English - Explicit emotion, speed, pitch and volume control - Multilingual synthesis and Sichuanese/Cantonese dialect control Recommended built-in voices: **UK_George**, **US_Emma**, **灵小柔** and **灵楷**. See the [voice catalog](https://github.com/inclusionAI/Ming-3.0-tts/blob/main/cookbooks/instruction_en.md#built-in-premium-voices) for the full list and voice descriptions. Use the exact `ip` values shown.

Ming-3.0-tts architecture

## Quick Start ### Installation ```bash git clone https://github.com/inclusionAI/Ming-3.0-tts.git cd Ming-3.0-tts pip install -r requirements.txt ``` Tested on NVIDIA H800-80GB and H20-96G with CUDA 12.4. `flash-attn` is optional; inference falls back to PyTorch SDPA. Numbers, dates, units and symbols are normalized with FST grammars. Install `pynini`, or pass `--no_text_normalization` to skip text normalization: ```bash conda install -c conda-forge pynini==2.1.6 ``` ### Download ```bash pip install "huggingface_hub[cli]" hf download inclusionAI/Ming-3.0-tts-0.5B --local-dir Ming-3.0-tts-0.5b ``` ### CLI Inference ```bash 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 ``` ### Python API ```python from ming_tts import MingTTS model = MingTTS("inclusionAI/Ming-3.0-tts-0.5B") model.synthesize( task="TTS", language_id="Chinese", # English for English text 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=True, output_path="output/tts.wav", ) ``` Supported tasks are `TTS`, `FREEFORM`, `VOICE_DESIGN`, `MULTILINGUAL`, `BASIC`, `EMOTION`, `DIALECT` and `IP`. See the [task cookbook](https://github.com/inclusionAI/Ming-3.0-tts/blob/main/cookbooks/instructions.md) for controls and examples. `MULTILINGUAL` requires `language_id` for German, Spanish, French, Korean or Japanese; `DIALECT` takes no language ID. Freeform uses its own task and cannot be combined with voice design, a built-in voice, emotion, dialect or attribute controls. With a reference clip it follows that voice; without a reference, the instruction can choose the voice. ### Verify the Checkpoint ```bash python tools/preflight.py --model_path /path/to/Ming-3.0-tts-0.5b ``` ## Performance Benchmark comparisons from the [project README](https://github.com/inclusionAI/Ming-3.0-tts#performance). **Bold** and underline mark the best and second-best values among the listed open-source models; ties share a mark. Seed-TTS and NVbench-GT are unranked reference rows. ### Zero-shot TTS (Seed-TTS testset)
Modelseed-tts-eval-zhseed-tts-eval-en
WER ↓SIM ↑WER ↓SIM ↑
Seed-TTS1.110%0.7962.240%0.762
GLM-TTS1.030%0.7612.230%0.672
Qwen3-TTS1.070%—1.390%—
Qwen3-TTS-12Hz-0.6B-Base0.920%—1.320%—
Qwen3-TTS-12Hz-1.7B-Base0.770%—1.240%—
CosyVoice3-0.5B1.160%0.7802.020%0.718
CosyVoice3-1.5B0.710%0.7751.450%0.695
F5-TTS1.560%0.7411.830%0.647
Ming-3.0-tts-0.5b0.800%0.7581.584%0.691
Ming-3.0-tts-tiny0.820%0.7811.666%0.728
### Multilingual (MiniMax test set, selected languages) Evaluated on the German, Spanish, French, Korean, and Japanese subsets of [MiniMaxAI/TTS-Multilingual-Test-Set](https://huggingface.co/datasets/MiniMaxAI/TTS-Multilingual-Test-Set).
ModelGermanSpanishFrenchKoreanJapanese
WER ↓SIM ↑WER ↓SIM ↑WER ↓SIM ↑CER ↓SIM ↑CER ↓SIM ↑
dots.tts0.452%0.7982.593%0.83919.878%0.7894.837%0.8535.134%0.838
VoxCPM20.963%0.7901.324%0.8363.380%0.7613.547%0.8394.358%0.825
Fish Audio S21.045%0.7141.216%0.7483.636%0.6981.675%0.7613.427%0.755
FireRedTTS30.348%0.8251.086%0.8684.366%0.8071.644%0.8643.418%0.825
Ming-3.0-tts-0.5b0.365%0.7531.099%0.7634.264%0.6751.728%0.7933.765%0.765
Ming-3.0-tts-tiny0.325%0.7910.971%0.7973.470%0.7336.000%0.7973.427%0.780

Each open-source model was evaluated using its recommended inference settings under a unified ASR-based WER/CER evaluation protocol.

### Paralinguistic Control (NVBench)
ModelSingleMulti
ChineseEnglishChineseEnglish
CERPCEROCERCERPCEROCERCERPCEROCERCERPCEROCER
NVbench-GT3.869.384.076.738.316.903.7923.715.077.2621.418.62
Orpheus-TTS11.3688.7713.919.0371.9210.6319.8384.8524.388.6871.4611.89
SMIIP-NV-CV28.8075.6411.3417.9256.8019.4710.6677.2014.7920.9354.4923.87
Emilia-NV-CV25.0540.006.6412.5055.3013.215.5448.748.0911.7160.2814.63
CosyVoice33.8557.695.867.8762.759.064.7561.948.266.3957.8410.69
NV-FlexiVoice6.9831.088.1511.8850.4313.218.2039.3710.399.6051.3213.76
NV-CV33.8027.694.908.3346.139.443.4430.044.846.7047.1310.10
Ming-3.0-tts-0.5b2.7022.803.207.6037.008.903.5036.608.107.9047.3010.40
Ming-3.0-tts-tiny2.8017.233.205.8632.957.054.4932.885.815.1246.137.68
### Dialect Control
ModelWSC-Eval-TTS-easyWSC-Eval-TTS-hardWSYue-TTS-eval-BaseWSYue-TTS-eval-Coverage
CER ↓SIM ↑ACC ↑CER ↓SIM ↑ACC ↑CER ↓SIM ↑ACC ↑CER ↓SIM ↑ACC ↑
Qwen3-TTS4.13%——7.35%————————
CosyVoice3-0.5B3.17%0.69676.05%4.07%0.72381.41%8.36%0.61191.70%8.95%0.65895.80%
CosyVoice3-1.5B + DiffRO-EMO7.14%0.703—9.06%0.601—14.38%0.812—13.74%0.826—
CosyVoice2-WSC4.28%0.728—8.87%0.626———————
CosyVoice2-WSC-SFT4.08%0.788—7.22%0.680———————
CosyVoice2-Yue——————10.33%0.821—9.49%0.834—
Sparks-TTS10.83%0.677—12.52%0.545—27.79%0.762—24.25%0.781—
Ming-3.0-tts-0.5b2.27%0.68980.68%3.35%0.69187.64%6.15%0.58698.00%8.15%0.62295.27%
Ming-3.0-tts-tiny2.53%0.72186.09%3.11%0.73888.94%7.40%0.60592.90%8.67%0.65395.16%
### Emotion Control (CV3)
ModelAverageText-RelatedText-Unrelated
HappySadAngryHappySadAngry
CosyVoice3-0.5B66.33%92%70%72%64%42%58%
CosyVoice3-1.5B63.00%86%64%72%64%44%48%
CosyVoice3-1.5B + DiffRO-EMO77.67%98%68%84%98%50%68%
F5-TTS64.67%92%52%72%80%28%64%
Sparks-TTS55.33%80%56%50%50%60%36%
Ming-3.0-tts-0.5b72.33%96%76%94%52%36%80%
Ming-3.0-tts-tiny74.00%98%82%90%60%36%78%
### Emotion Control (CV3N, hard emotion control)
ModelAverageText-RelatedText-Unrelated
HappySadAngryHappySadAngry
CosyVoice3-0.5B40.00%68%30%78%14%4%46%
Ming-3.0-tts-0.5b37.33%70%36%70%14%4%30%
Ming-3.0-tts-tiny30.67%60%40%44%18%10%12%
### Voice Design (InstructTTSEval-ZH)
ModelAverageAPSDSDRP
Qwen3-TTS77.13%85.20%81.10%65.10%
Mimo-Audio-7B-Instruct70.50%75.70%74.30%61.50%
VoiceSculptor67.30%75.70%64.70%61.50%
VoxInstruct47.47%47.50%52.30%42.60%
Ming-3.0-tts-0.5b72.29%78.54%77.28%61.06%
Ming-3.0-tts-tiny75.37%85.10%79.90%61.10%
### Basic Attribute Control
ModelSpeedVolumePitchAverageWER ↓SIM ↑
CosyVoice3-0.5B100.00%97.67%65.33%87.67%1.210%0.580
Ming-3.0-tts-0.5b99.00%95.00%86.00%93.30%0.317%0.712
Ming-3.0-tts-tiny93.00%93.00%78.70%88.23%0.274%0.742
## Limitations - Speaker similarity is weaker than content accuracy; French SIM is the clearest multilingual gap. - NVBench Multi trails Single; stability with several paralinguistic events in one sentence is still improving. - Emotion control is harder when the words carry no emotional signal. - Pitch is less accurate than speed and volume, and open-ended Voice Design does not yet match the strongest evaluated system. ## Usage Disclaimer - Voice cloning and built-in character voices are provided for research and legitimate creative use. Do not use them to impersonate a real person without consent, to deceive, or for any illegal purpose. - Label synthetic speech as synthetic where it could be mistaken for a real recording. - The authors accept no liability for misuse. If you find abuse of this project, please report it to us. ## Acknowledgements - [Qwen2.5](https://github.com/QwenLM/Qwen2.5) for the backbone - [CAM++](https://modelscope.cn/models/iic/speech_campplus_sv_zh-cn_16k-common) for speaker embedding extraction - [Ming-omni-tts](https://github.com/inclusionAI/Ming-omni-tts), the previous generation, whose audio tokenizer this release builds on ## License Released under the [MIT license](https://github.com/inclusionAI/Ming-3.0-tts/blob/main/LICENSE).