Ming-3.0 Speech Series Open-Sourced: 5 Models for TTS, Audio Understanding & ASR

Overview

Ant Group's Ling team has open-sourced the Ming-3.0 speech series , comprising five models across three capabilities: speech synthesis (Ming-3.0-tts-0.5B, Ming-3.0-tts-tiny), audio understanding and dialogue (Ming-3.0-audio-flash, Ming-3.0-audio-tiny), and speech recognition (Ming-3.0-asr-tiny). Developers can pick models per task or combine them into custom voice workflows.

1. Speech Synthesis: Ming-3.0-tts Series

Voice Design & Director Mode

Ming-3.0-tts supports Voice Design via textual description of target timbre or reference audio. A distinctive director mode lets creators specify performance instructions — emotion shifts, word-level stress, and rhythm pacing — in plain language. For example, a director prompt "Sound angry but controlled. Stress the accusation and make the final question sharp, without shouting." applied to the text "You changed the plan without telling anyone. How were we supposed to know what to do?" yields a controlled-anger delivery.

Paralinguistic Tags

Inline tags such as [Laughter], [Sigh], [Cough] can be inserted in the synthesis text to place non-verbal expressions precisely. Example: "他们一进门就笑成一团，[Laughter]整个屋子都热闹起来了。"

Benchmark Results (Artificial Analysis, Oct 2026)

Pronunciation Robustness (proper nouns, abbreviations, numbers): 93.2% , tied for #1 on the leaderboard.

Controlled Voice Elo: 1069 , #1 among open-weight models.

Provider Voices Elo: 1188 , #3 among open-weight models.

Model size: 0.5B parameters (Ming-3.0-tts-0.5B) achieves these results.

2. Audio Understanding & Dialogue: Ming-3.0-audio Series

Long-Audio Processing Architecture

Ming-3.0-audio-flash/tiny accept long audio directly for transcription, summarization, and viewpoint analysis with structured output. Unlike traditional cascaded ASR-then-analyze pipelines, they preserve emotion, speaker identity, and environmental cues. Key architectural choices:

Input audio representation at 6.25 Hz low frame rate , shortening sequence length for equal duration.

Language backbone based on Ling-3.0 .

Hybrid attention: KDA + Gated MLA .

Sparse Mixture-of-Experts (MoE) for efficient long-context inference.

Benchmark Scores

MMAU (comprehensive audio understanding): 82.00

VoiceBench (speech QA & instruction following): 86.74

BigBenchAudio (speech reasoning, S2T): 99.40 , #1 among listed models.

SpeechFCEval (speech tool calling): 82.60

3. Speech Recognition: Ming-3.0-asr-tiny

Context-Aware ASR with Hotwords & Dialogue History

Ming-3.0-asr-tiny (8.6B A2B) tackles dialect accents, hotword recognition, and dialogue context injection. It leverages high-quality corpora (Mandarin, English, dialects, domain hotwords) and the LLM backbone's contextual understanding. Two types of scene information can be supplied at inference:

Hotword list : product names, person names, place names, industry terms.

Dialogue context : preceding conversation turns to disambiguate homophones and rare names.

No extra model training required. Deployed in Ant Group scenarios: 阿福医生助手, 支付宝搜索, 蚁小保 real-time voice dialogue, safety verification.

Example: Cantonese Navigation with Hotwords

Hotwords provided: 佐敦道、弥敦道、波鞋街、西洋菜街 . Input: "弥敦道怎么走?" → correctly recognized.

Accuracy Benchmarks

Average word error rate across four test sets (domain hotwords, dialect, Mandarin, English): 4.41% . Outperforms closed-source Fun-ASR 1.5 and open-source FireRedASR2-LLM (8B), Qwen3-ASR-1.7B.

Limitations & Future Work

Focus on enabling Voice Agent for long-horizon complex tasks: multi-turn dialogue to understand needs, initiate tasks, track progress, retain context and user preferences.

Access & Resources

Online demo: https://ling.tbox.cn (Ling Studio Chat, select Ming-3.0-audio-flash).

API service: Ling Studio provides APIs for Ming-3.0-tts-0.5B, Ming-3.0-audio-flash, Ming-3.0-asr-tiny.

Model hubs (Hugging Face & ModelScope):

Ming-3.0-tts-0.5B: https://huggingface.co/inclusionAI/Ming-3.0-tts-0.5B,

https://www.modelscope.cn/models/inclusionAI/Ming-3.0-tts-0.5B

Ming-3.0-tts-tiny: https://huggingface.co/inclusionAI/Ming-3.0-tts-tiny,

https://www.modelscope.cn/models/inclusionAI/Ming-3.0-tts-tiny

Ming-3.0-audio-flash: https://huggingface.co/inclusionAI/Ming-3.0-audio-flash,

https://www.modelscope.cn/models/inclusionAI/Ming-3.0-audio-flash

Ming-3.0-audio-tiny: https://huggingface.co/inclusionAI/Ming-3.0-audio-tiny,

https://www.modelscope.cn/models/inclusionAI/Ming-3.0-audio-tiny

Ming-3.0-asr-tiny: https://huggingface.co/inclusionAI/Ming-3.0-asr-tiny,

https://www.modelscope.cn/models/inclusionAI/Ming-3.0-asr-tiny