Ming-3.0-tts-tiny

Image: Ming-3.0-tts Logo

Ming-3.0-tts-tiny#

7.9B-A1.3B multilingual MoE speech synthesis with paralinguistic and free-form control.

Image: image GitHub | Image: image Blog & Demos | Image: image ModelScope

This repository contains Ming-3.0-tts-tiny, the Ming-3.0-tts mixture-of-experts checkpoint built on Ling V3. Its Ling V3 backbone has 7.9B total parameters and activates about 1.3B parameters per token. The repository includes sharded weights, supporting model code, audio tokenizer, speaker encoder and text tokenizer.

Inference requires the Ming-3.0-tts GitHub repository. 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#

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
• Multilingual synthesis and Sichuanese/Cantonese dialect control
• 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
• Multi-speaker dialogue from a script and speaker references (podcast)

Recommended built-in voices: UK_George, US_Emma, 灵小柔 and 灵楷. See the voice catalog for the full list and voice descriptions. Use the exact ip values shown.

Image: Ming-3.0-tts-tiny architecture

Quick Start#

Installation#

git clone https://github.com/inclusionAI/Ming-3.0-tts.git
cd Ming-3.0-tts
conda create -n ming-tiny python=3.10.15 -y
conda activate ming-tiny
pip install -r requirements-tiny.txt

Tiny uses a separate reference inference environment: Python 3.10.15, PyTorch and torchaudio 2.7.1+cu118 (CUDA 11.8 wheels), and transformers==4.46.3. Install the pinned dependencies from requirements-tiny.txt in a dedicated environment, as other dependency versions differ from 0.5B. We recommend using NVIDIA H800 for inference.

Download#

pip install "huggingface_hub[cli]"
hf download inclusionAI/Ming-3.0-tts-tiny --local-dir Ming-3.0-tts-tiny

CLI Inference#

python infer_tiny.py \
  --model_path inclusionAI/Ming-3.0-tts-tiny \
  --task dialect \
  --text "今日天气几好，我哋食完饭一齐去海边行下啦。" \
  --dialect 广粤话 \
  --prompt_wav data/wavs/10002287-00000094.wav \
  --output_path output/tiny.wav

Python API#

Loading configures the release runtime (bf16, portable MoE loop, fp32 recurrent state) and checks the checkpoint against it; generate() defaults to cfg_scale=2.0, temperature=0, flow_steps=10, and sigma=0.25.

from ming_tiny import MingTinyTTS

tts = MingTinyTTS.from_pretrained("inclusionAI/Ming-3.0-tts-tiny", device="cuda:0")
tts.generate(
    task="seed",
    text="傍晚的风从山口吹过来，远处的云慢慢染成了金色。",
    prompt_wav="data/wavs/10002287-00000094.wav",
    prompt_text="在此奉劝大家别乱打美白针。",
    target_language="Chinese",          # English for English text
    output_path="outputs/seed.wav",
)

Supported Tasks#

Each call selects one task and accepts the inputs listed below:

Freeform always requires a reference clip and cannot be combined with voice design, a built-in voice, emotion, dialect or attribute controls. For podcast, start every dialogue line with speaker_N: and provide one reference audio clip and transcript per speaker, in speaker order. Repeat --prompt_wav and --prompt_text on the command line.

See the Tiny examples for a Podcast Python example and examples/one_click/ for scripts for every task. python infer_tiny.py --help lists the CLI arguments.

Verify the Checkpoint#

python tools/preflight_tiny.py --model_path /path/to/Ming-3.0-tts-tiny

Serving with vLLM-Omni#

For installation, follow the vLLM-Omni official installation guide; vLLM-Omni serves this checkpoint through the OpenAI-compatible /v1/audio/speech endpoint. The current recipe targets vLLM 0.32.0. Use a separate environment from the reference inference setup above.

Start the server in one terminal:

vllm serve inclusionAI/Ming-3.0-tts-tiny --omni --port 8000

The HF model ID selects the two-stage Ming3-TTS pipeline automatically and downloads the weights and required assets on first launch. Both stages share GPU 0. A complete local model directory may also be used; pass --revision <HF-commit> to pin a remote checkpoint.

After the server is ready, run this reference-free voice design example in another terminal:

curl --fail-with-body --silent --show-error http://127.0.0.1:8000/v1/audio/speech \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "inclusionAI/Ming-3.0-tts-tiny",
    "input": "傍晚的风从山口吹过来，远处的云慢慢染成了金色。",
    "instructions": "{\"风格\":\"年轻女声，声音清晰自然。\"}",
    "response_format": "wav",
    "extra_params": {"temperature": 0.0, "ming3_cfg": 1.0},
    "seed": 1234
  }' --output ming3_tiny_design.wav

The output is a mono 44.1 kHz WAV. Voice cloning uses ref_audio and its matching transcript in ref_text; speech controls use instructions. For voice cloning, dialect and emotion controls, built-in voices, multi-speaker Podcast, and PCM streaming, see the upstream Ming-3.0-TTS recipe.

Serving with SGLang-Omni#

SGLang-Omni serves this checkpoint on one GPU:

sgl-omni serve --model-path inclusionAI/Ming-3.0-tts-tiny

Synthesize speech through the OpenAI-compatible endpoint:

curl -X POST http://127.0.0.1:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "inclusionAI/Ming-3.0-tts-tiny",
    "input": "Get the trust fund to the bank early.",
    "response_format": "wav",
    "seed": 123
  }' \
  --output speech.wav

response_format accepts wav, mp3, flac, and pcm. For zero-shot voice cloning, pass ref_audio (a server-side path or a data: URL) together with ref_text, the transcript of the reference recording. See the SGLang-Omni cookbook for details.

Performance#

Benchmark comparisons from the project README.

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)#

| Model | seed-tts-eval-zh | seed-tts-eval-en

| WER ↓ | SIM ↑ | WER ↓ | SIM ↑

| Seed-TTS | 1.110% | 0.796 | 2.240% | 0.762

| GLM-TTS | 1.030% | 0.761 | 2.230% | 0.672

| Qwen3-TTS | 1.070% | — | 1.390% | —

| Qwen3-TTS-12Hz-0.6B-Base | 0.920% | — | 1.320% | —

| Qwen3-TTS-12Hz-1.7B-Base | 0.770% | — | 1.240% | —

| CosyVoice3-0.5B | 1.160% | 0.780 | 2.020% | 0.718

| CosyVoice3-1.5B | 0.710% | 0.775 | 1.450% | 0.695

| F5-TTS | 1.560% | 0.741 | 1.830% | 0.647

| Ming-3.0-tts-0.5b | 0.800% | 0.758 | 1.584% | 0.691

| Ming-3.0-tts-tiny | 0.820% | 0.781 | 1.666% | 0.728

Multilingual (MiniMax test set, selected languages)#

Evaluated on the German, Spanish, French, Korean, and Japanese subsets of MiniMaxAI/TTS-Multilingual-Test-Set.

| Model | German | Spanish | French | Korean | Japanese

| WER ↓ | SIM ↑ | WER ↓ | SIM ↑ | WER ↓ | SIM ↑ | CER ↓ | SIM ↑ | CER ↓ | SIM ↑

| dots.tts | 0.452% | 0.798 | 2.593% | 0.839 | 19.878% | 0.789 | 4.837% | 0.853 | 5.134% | 0.838

| VoxCPM2 | 0.963% | 0.790 | 1.324% | 0.836 | 3.380% | 0.761 | 3.547% | 0.839 | 4.358% | 0.825

| Fish Audio S2 | 1.045% | 0.714 | 1.216% | 0.748 | 3.636% | 0.698 | 1.675% | 0.761 | 3.427% | 0.755

| FireRedTTS3 | 0.348% | 0.825 | 1.086% | 0.868 | 4.366% | 0.807 | 1.644% | 0.864 | 3.418% | 0.825

| Ming-3.0-tts-0.5b | 0.365% | 0.753 | 1.099% | 0.763 | 4.264% | 0.675 | 1.728% | 0.793 | 3.765% | 0.765

| Ming-3.0-tts-tiny | 0.325% | 0.791 | 0.971% | 0.797 | 3.470% | 0.733 | 6.000% | 0.797 | 3.427% | 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)#

| Model | Single | Multi

| Chinese | English | Chinese | English

| CER | PCER | OCER | CER | PCER | OCER | CER | PCER | OCER | CER | PCER | OCER

| NVbench-GT | 3.86 | 9.38 | 4.07 | 6.73 | 8.31 | 6.90 | 3.79 | 23.71 | 5.07 | 7.26 | 21.41 | 8.62

| Orpheus-TTS | 11.36 | 88.77 | 13.91 | 9.03 | 71.92 | 10.63 | 19.83 | 84.85 | 24.38 | 8.68 | 71.46 | 11.89

| SMIIP-NV-CV2 | 8.80 | 75.64 | 11.34 | 17.92 | 56.80 | 19.47 | 10.66 | 77.20 | 14.79 | 20.93 | 54.49 | 23.87

| Emilia-NV-CV2 | 5.05 | 40.00 | 6.64 | 12.50 | 55.30 | 13.21 | 5.54 | 48.74 | 8.09 | 11.71 | 60.28 | 14.63

| CosyVoice3 | 3.85 | 57.69 | 5.86 | 7.87 | 62.75 | 9.06 | 4.75 | 61.94 | 8.26 | 6.39 | 57.84 | 10.69

| NV-FlexiVoice | 6.98 | 31.08 | 8.15 | 11.88 | 50.43 | 13.21 | 8.20 | 39.37 | 10.39 | 9.60 | 51.32 | 13.76

| NV-CV3 | 3.80 | 27.69 | 4.90 | 8.33 | 46.13 | 9.44 | 3.44 | 30.04 | 4.84 | 6.70 | 47.13 | 10.10

| Ming-3.0-tts-0.5b | 2.70 | 22.80 | 3.20 | 7.60 | 37.00 | 8.90 | 3.50 | 36.60 | 8.10 | 7.90 | 47.30 | 10.40

| Ming-3.0-tts-tiny | 2.80 | 17.23 | 3.20 | 5.86 | 32.95 | 7.05 | 4.49 | 32.88 | 5.81 | 5.12 | 46.13 | 7.68

Dialect Control#

| Model | WSC-Eval-TTS-easy | WSC-Eval-TTS-hard | WSYue-TTS-eval-Base | WSYue-TTS-eval-Coverage

| CER ↓ | SIM ↑ | ACC ↑ | CER ↓ | SIM ↑ | ACC ↑ | CER ↓ | SIM ↑ | ACC ↑ | CER ↓ | SIM ↑ | ACC ↑

| Qwen3-TTS | 4.13% | — | — | 7.35% | — | — | — | — | — | — | — | —

| CosyVoice3-0.5B | 3.17% | 0.696 | 76.05% | 4.07% | 0.723 | 81.41% | 8.36% | 0.611 | 91.70% | 8.95% | 0.658 | 95.80%

| CosyVoice3-1.5B + DiffRO-EMO | 7.14% | 0.703 | — | 9.06% | 0.601 | — | 14.38% | 0.812 | — | 13.74% | 0.826 | —

| CosyVoice2-WSC | 4.28% | 0.728 | — | 8.87% | 0.626 | — | — | — | — | — | — | —

| CosyVoice2-WSC-SFT | 4.08% | 0.788 | — | 7.22% | 0.680 | — | — | — | — | — | — | —

| CosyVoice2-Yue | — | — | — | — | — | — | 10.33% | 0.821 | — | 9.49% | 0.834 | —

| Sparks-TTS | 10.83% | 0.677 | — | 12.52% | 0.545 | — | 27.79% | 0.762 | — | 24.25% | 0.781 | —

| Ming-3.0-tts-0.5b | 2.27% | 0.689 | 80.68% | 3.35% | 0.691 | 87.64% | 6.15% | 0.586 | 98.00% | 8.15% | 0.622 | 95.27%

| Ming-3.0-tts-tiny | 2.53% | 0.721 | 86.09% | 3.11% | 0.738 | 88.94% | 7.40% | 0.605 | 92.90% | 8.67% | 0.653 | 95.16%

Emotion Control (CV3)#

| Model | Average | Text-Related | Text-Unrelated

| Happy | Sad | Angry | Happy | Sad | Angry

| CosyVoice3-0.5B | 66.33% | 92% | 70% | 72% | 64% | 42% | 58%

| CosyVoice3-1.5B | 63.00% | 86% | 64% | 72% | 64% | 44% | 48%

| CosyVoice3-1.5B + DiffRO-EMO | 77.67% | 98% | 68% | 84% | 98% | 50% | 68%

| F5-TTS | 64.67% | 92% | 52% | 72% | 80% | 28% | 64%

| Sparks-TTS | 55.33% | 80% | 56% | 50% | 50% | 60% | 36%

| Ming-3.0-tts-0.5b | 72.33% | 96% | 76% | 94% | 52% | 36% | 80%

| Ming-3.0-tts-tiny | 74.00% | 98% | 82% | 90% | 60% | 36% | 78%

Emotion Control (CV3N, hard emotion control)#

| Model | Average | Text-Related | Text-Unrelated

| Happy | Sad | Angry | Happy | Sad | Angry

| CosyVoice3-0.5B | 40.00% | 68% | 30% | 78% | 14% | 4% | 46%

| Ming-3.0-tts-0.5b | 37.33% | 70% | 36% | 70% | 14% | 4% | 30%

| Ming-3.0-tts-tiny | 30.67% | 60% | 40% | 44% | 18% | 10% | 12%

Voice Design (InstructTTSEval-ZH)#

| Model | Average | APS | DSD | RP

| Qwen3-TTS | 77.13% | 85.20% | 81.10% | 65.10%

| Mimo-Audio-7B-Instruct | 70.50% | 75.70% | 74.30% | 61.50%

| VoiceSculptor | 67.30% | 75.70% | 64.70% | 61.50%

| VoxInstruct | 47.47% | 47.50% | 52.30% | 42.60%

| Ming-3.0-tts-0.5b | 72.29% | 78.54% | 77.28% | 61.06%

| Ming-3.0-tts-tiny | 75.37% | 85.10% | 79.90% | 61.10%

Basic Attribute Control#

| Model | Speed | Volume | Pitch | Average | WER ↓ | SIM ↑

| CosyVoice3-0.5B | 100.00% | 97.67% | 65.33% | 87.67% | 1.210% | 0.580

| Ming-3.0-tts-0.5b | 99.00% | 95.00% | 86.00% | 93.30% | 0.317% | 0.712

| Ming-3.0-tts-tiny | 93.00% | 93.00% | 78.70% | 88.23% | 0.274% | 0.742

Podcast#

| Model | CER ↓ | cpSIM ↑ | UTMOS ↑

| FireRedTTS2 | 3.34% | 0.512 | 1.90

| SoulX-Podcast | 2.20% | 0.599 | 2.09

| VibeVoice-1.5B | 12.87% | 0.455 | 1.74

| MOSS-TTSD | 8.62% | 0.421 | 1.70

| ZipVoice-Dia | 3.39% | 0.553 | 2.24

| MoonCast | 27.43% | 0.441 | 1.76

| Ming-3.0-tts-tiny | 2.03% | 0.561 | 2.12

Limitations#

• Korean CER (6.000%) is the clearest multilingual content-accuracy gap for this checkpoint.
• Speaker similarity remains weaker than content accuracy on several multilingual and dialect evaluations.
• NVBench Multi trails Single on several metrics; stability with multiple 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.
• Podcast audio quality remains a limitation: UTMOS is 2.12, below ZipVoice-Dia's 2.24 and the historical Ming-omni-tts-0.5B score of 2.25.

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#

• Ling V3 for the mixture-of-experts backbone

The publisher's README is longer; it is cut here. Read the rest at the source.