On October 20, Ant Group's Bailian open-sourced the Ming-3.0 speech series, providing three series totaling five models for three types of tasks—voice creation, long-form speech analysis, and business speech transcription: Ming-3.0-tts (0.5B, tiny), Ming-3.0-audio (flash, tiny), and Ming-3.0-asr-tiny. The weights are available on Hugging Face and ModelScope, and Ling Studio provides APIs and an online experience.
On the synthesis side, Ming-3.0-tts's Voice Design can describe a target voice or reference audio with text, and Director Mode allows using text to specify emotional changes, emphasized words, and pacing, with support for inline paralinguistic tags marking the positions of laughter, sighs, and coughs. The series has been deployed in the Afu APP's voice calls, doctor voice cloning, and Xiaohu smart office.
On the understanding side, Ming-3.0-audio-flash and tiny can directly transcribe, summarize, and analyze viewpoints in long audio with structured output, preserving emotion, speaker, and environmental information; their input side uses 6.25Hz low-frame-rate audio representations to shorten sequences, and the language backbone is based on Ling-3.0, combining KDA with Gated MLA hybrid attention and sparse MoE.
On the recognition side, Ming-3.0-asr-tiny's Context ASR can accept hotword lists and dialogue context to help disambiguate homophones without additional training, and has been used in Afu Doctor Assistant, Alipay Search, and security identity verification.
In terms of real-world testing, in Artificial Analysis' 2026-10 evaluation, Ming-3.0-tts-0.5B scored 93.2% on Pronunciation Robustness, tied for first place across the entire leaderboard, and ranked first among open-weight models on Controlled Voice with 1069 Elo.
Ming-3.0-audio-flash scored 82.00 and 86.74 on MMAU and VoiceBench respectively, scored 99.40 on BigBenchAudio (S2T) to rank first, and scored 82.60 on SpeechFCEval. Ming-3.0-asr-tiny (8.6B A2B) achieved an average error rate of 4.41% across the four types of evaluations, outperforming Fun-ASR1.5, FireRedASR2-LLM (8B), and Qwen3-ASR-1.7B.