Ming-3.0 : Ant Group launches five open models for voice generation, understanding and transcription

Ant Group releases five open-weight Ming-3.0 models for text-to-speech, audio understanding, and transcription, with emotion control.

Creating a voice from a description, making it express an emotion, analyzing an interview, or transcribing a conversation despite accents and background noise. With Ming-3.0, the Ant Ling team from Ant Group releases five open-weights models covering three areas of speech processing: synthesis, audio understanding, and automatic speech recognition. This new family combines a compact 500-million-parameter model with mixture-of-experts architectures reaching up to 126 billion parameters in total. The weights are available on Hugging Face and ModelScope, with planned integrations for the vLLM and SGLang inference frameworks. Voices that can be described and directed in writing

The first branch, Ming-3.0-tts, focuses on speech generation from text. It includes two models: a 500-million-parameter version and Ming-3.0-tts-tiny, built on a 7.9-billion-parameter Mixture-of-Experts architecture, of which about 1.3 billion are activated per token. This distinction makes it possible to offer configurations tailored to different generation and deployment needs.

One of the highlighted features is Voice Design. Instead of selecting only a pre-recorded voice, the user can describe its characteristics using natural language prompts. The model then attempts to produce a timbre and performance matching that description. It is also possible to provide a short audio clip to replicate the characteristics of a reference voice, without additional specific fine-tuning.

Control is not limited to voice identity. Both models accept instructions regarding emotion, speech rate, pitch, and volume, as well as indications embedded directly into the text. Fourteen tags make it possible to request a laugh, a sigh, a hesitation, a cough, or specific vocal reactions. These functions target storytelling, virtual characters, and conversational assistants, where intonation can play a role as important as the words spoken.

Synthesis supports seven languages: Chinese, English, German, Spanish, French, Korean, and Japanese, along with control options for certain Chinese dialects, including Cantonese and Sichuanese. The TTS Tiny variant also offers multi-speaker dialogue generation from a script and voice references, with potential applications in podcasts or long-form narrative formats. Reported results among top open voice models

Ant Ling particularly highlights the performance of Ming-3.0-tts-0.5B. According to the announcement, the model achieves 1,069 Elo points in the Controlled Voice ranking by Artificial Analysis, placing it at the top spot among open-weights models in this category. It also scores 1,188 Elo points in the Provider Voice evaluation, reported in third place, and achieves a 93.2% score in pronunciation robustness, tied with the top-evaluated models.

These figures relate to distinct benchmarks, not a single unified metric of voice quality. The Controlled Voice ranking specifically evaluates the ability to produce instruction-directed speech, whereas pronunciation robustness focuses on adherence to the spoken text. A model can therefore perform well in one category without achieving identical results in speaker fidelity, long-dialogue stability, or emotional expression.

Detailed evaluations published alongside the weights also reveal trade-offs between the two variants. On multilingual tests reproduced by the team, Ming-3.0-tts-tiny achieves a word error rate of 0.971% in Spanish and 3.470% in French, compared to 1.099% and 4.264% for the 0.5B version. Conversely, the authors report that similarity to the reference speaker remains less convincing in certain languages, particularly French for the compact version.

Simultaneously mastering multiple vocal events and the nuance of certain emotions also remain areas for improvement. The reported performance should therefore be understood as results achieved on targeted benchmarks, rather than a guarantee of flawless expressive delivery in every scenario. Understanding an audio recording rather than merely transcribing it

The second branch, Ming-3.0-audio, goes beyond recognizing spoken words. It is designed to interpret audio content and answer questions about it. Its applications include meeting analysis, interviews, environmental sound or music understanding, as well as services capable of receiving voice instructions.

The largest model, Ming-3.0-audio-flash, relies on a mixture-of-experts architecture totaling 126 billion parameters, with 6.5 billion active per inference step. Its audio-tiny variant features 9.3 billion parameters, with 2.7 billion active. The total size of these architectures does not correspond to the parameter volume deployed during each operation. This design aims to combine high representation capacity with more contained compute costs than full parameter activation.

Both models accept audio content accompanied by text instructions. They can notably identify information within a recording, answer an orally formulated question, or produce text output paired, depending on the configuration, with a synthetic voice response. Ming-3.0-audio-flash can also generate tool-use requests that external applications are then responsible for executing.

This latter capability brings the model closer to voice assistants capable of acting on digital services. Understanding a request and formulating a tool call are handled by the model, but actual execution still depends on the integrating software and the permissions granted to it. Technical documentation indicates that primary use cases currently target Chinese and English, across various accents and dialects. Transcription that accounts for context

The third branch consists of Ming-3.0-asr-tiny, a model specialized in automatic speech recognition. It contains approximately 8.6 billion parameters, with 2 billion active, and targets call transcription, voice search, and environments involving technical terms or domain-specific jargon.

Its operation includes a feature called Context ASR. Rather than requiring developers