Ming-3.0-asr
an open‑source locally‑deployable end‑to‑end noise‑robust ASR model from Ant inclusionAI with context hot‑word enhancement for Mandarin, English and 14 Chinese dialects.
Note: Please open in the PC Client. Go to download >>>
Features
Supported Languages
Dialects
System Requirements
Minimum 16GB RAM. 26GB+ storage recommended.
macOS 15+: Supports M-series chips.
Windows 10/11: NVIDIA GPU with 8GB+ VRAM recommended.
Note: For NVIDIA GPUs, install a newer driver.Introduction
Ming‑3.0‑asr (Ming‑3.0‑asr‑tiny) is an open‑source Automatic Speech Recognition (ASR, speech‑to‑text) model suite developed by inclusionAI (Ant Ling), Ant Group, released under the MIT license. Its inference code, model weights and evaluation reports are fully open for offline local deployment. Built with an 8.6B A2B sparse architecture, it has been validated in large‑scale industrial scenarios inside Ant Group. It balances recognition accuracy under noisy conditions and local inference performance. It can work standalone or pair with Ming‑3.0‑tts from the same series to build a complete speech‑dialog pipeline. It has been deployed in internal Ant services including medical voice assistants, Alipay voice search, real‑time voice conversations and identity verification systems.
Core Functions & Features
- Noise‑robust recognition for challenging acoustic conditions Optimized for real‑world recording scenarios. It handles background noise, room reverberation, far‑field capture and phone‑call audio, maintaining good accuracy outside quiet studio environments. It supports both offline file transcription and streaming real‑time recognition.
- Context‑aware hot‑word enhancement (highlight feature) No fine‑tuning required. During inference, you can pass custom hot‑word lists and conversation history to improve recognition accuracy for proper nouns, personal names, product terms and domain‑specific vocabulary. It reduces recognition errors for vertical domains such as healthcare, government services, finance and customer service.
- Supported languages & dialects (based on official benchmark table)
✅ Primary languages: Mandarin Chinese, English
✅ 14 Chinese dialects verified by internal benchmark:
Cantonese, Chuanyu(Sichuanese), Hunan, Minnan, Shanghai, Anhui, Northeast, Henan, Hubei, Jiangsu, Hakka, Shaanxi, Shandong, Tianjin, Yunnan
Note: The official benchmark has evaluated all 14 dialects. Cantonese and Chuanyu(Sichuanese) achieve the best performance and are the officially prioritized dialects.
- Long‑audio transcription with timestamp output Supports one‑shot transcription for multi‑hour audio. Word‑level timestamps are available for subtitle creation, audio archiving and meeting minutes. It can process multi‑speaker recordings and segment utterances by different speakers.
- Local‑deployable for diverse hardware The sparse architecture reduces video‑memory consumption while keeping high recognition quality. It runs on consumer‑grade GPUs and supports large‑scale server deployment. Typical use‑cases include meeting transcription, subtitle generation, voice customer service, voice note‑taking and conversational bots.
- Strong performance in vertical domains It has been specially evaluated on five internal domain datasets: chat, government service, healthcare, knowledge and local life, delivering stable recognition results in professional scenarios.
Underlying Technology
It is built on inclusionAI’s self‑developed Ling V3 A2B sparse‑expert architecture, adopting an end‑to‑end ASR paradigm combining audio encoder and large‑language‑model backbone. Raw audio signals are encoded and fed into the sparse MoE LLM backbone. It fuses conversation context and user‑provided hot‑word prompts to produce transcribed text directly. The A2B sparse design activates only partial expert parameters, lowering compute cost while delivering noise‑resilient, context‑aware speech recognition.