Ming-3.0-tts Open-Source Text-to-Speech Model Family Released with 0.5B and Tiny MoE Variants

Ming-3.0-tts has been released as an open-source text-to-speech (TTS) model family featuring two core variants: a 0.5B dense model and a 7.9B-A1.3B Mixture of Experts (MoE) tiny model. The project provides inference code, evaluation benchmarks, and support for multilingual speech, voice cloning, voice design, and natural-language delivery control.

Model Architecture and Variants

Both models share a unified generation framework that combines a language model conditioned with text and control tokens, a DiT patch head, and an audio tokenizer:

• Ming-3.0-tts-0.5b: A dense model tested on hardware environments such as NVIDIA H800 and H20, utilizing PyTorch and optional flash-attention.
• Ming-3.0-tts-tiny: A MoE model built on Ling V3 (7.9B total, 1.3B active parameters), requiring a dedicated reference inference environment with Python 3.10.15 and pinned dependencies.

Core Capabilities and Control Features

Ming-3.0-tts offers distinct task modes designed for specific generation needs:

• Voice Creation: Supports zero-shot voice cloning from reference audio, voice design via text descriptions, and built-in character voices.
• Free-Form & Paralinguistic Control: Accepts natural-language instructions to guide delivery and includes 14 inline tags for expressions such as laughter, sighs, and breathing in Chinese and English.
• Multilingual & Dialects: Synthesizes speech in Chinese, English, German, Spanish, French, Korean, and Japanese, alongside Sichuanese and Cantonese dialects.
• Attributes & Dialogue: Allows adjustment of emotion, speed, pitch, and volume, with multi-speaker dialogue generation supported by the tiny variant.

Benchmark Performance and Evaluations

Independent evaluations and project benchmarks highlight varied strengths across metrics:

• Pronunciation & Paralinguistic: The models achieve strong results in pronunciation robustness and single-tag paralinguistic metrics on benchmarks like NVBench.
• Multilingual & Dialects: The tiny variant records competitive performance in German and Spanish word error rates, while the 0.5b variant excels in Cantonese base metrics.
• Limitations: Areas slated for future improvement include multilingual speaker similarity, tiny's Korean accuracy, text-unrelated emotion handling, and role-play instruction following.

Practical Implications and Verification

The release includes cookbook guides and one-click example scripts to help developers verify performance and integrate the models. Users must adhere to responsible AI guidelines: tools should not be used to impersonate individuals without consent, and synthetic speech should be properly labeled.

To verify the original announcement, review the official GitHub repository for codebases, environment configuration files, and benchmark datasets.