Trending
Hot topicsCalmed down

Ant Group’s inclusionAI open-sources five Ming-3.0 speech models

12 Articles5 Reporting sourcesUpdated 15 days ago

Learn about this first

Ant Ling (inclusionAI) has open-sourced the Ming-3.0 speech series, comprising five models covering three types of tasks: speech synthesis, audio understanding and dialogue, and speech recognition. Speech synthesis includes the 0.5B Ming-3.0-tts and a Tiny long-form speech model with 7.9B total parameters and 1.3B active parameters; audio understanding and dialogue include Flash (126B total parameters, 6.5B active) and Tiny (9.3B total parameters, 2.7B active); and speech recognition is handled by the Ming-3.0-asr-tiny, with 8.6B total parameters and 2B active parameters. According to Ant inclusionAI’s GitHub repository, Ming-3.0-tts is open-sourced under the MIT license, with weights available for download from Hugging Face and ModelScope.

Third-party evaluation organization Artificial Analysis says Ming-3.0-tts has become the highest-ranked open-weight model in its Controlled Voice TTS Arena. It also topped the pronunciation robustness leaderboard with a score of 93.2%, tying with Fish Audio Drama 3 Preview and exceeding Eleven v4’s 91.7%. The organization further said that the 0.5B version is the only model among the top three on the relevant TTS leaderboard with fewer than 1B parameters.

For deployment, both vLLM-Omni and SGLang-Omni provide day-one support. One developer ran Ming-3.0-tts-0.5B locally on an RTX 3090, reporting memory usage of about 3.3GB and disk usage of under 3GB, and completed voice cloning with a five-second recording sample of their own voice.

AI-generated from reports · updated 15 days ago

Latest developments

Ant Ling Ming-3.0 voice cloning: local hands-on test

Event progress8

  1. User shares local test of Ming-3.0-tts

    Ant Ling: Ant Ling Ming-3.0 voice cloning: local hands-on test
  2. Ant Group Ming-3.0-tts TTS leaderboard performance

    Artificial Analysis: Ant Group's Ming-3.0-tts enters the top three of the Artificial Analysis TTS leaderboard with 0.5B parameters
  3. Ant Group's Bailling thanks you for sharing the Ming-3.0 speech model

    inclusionAI: inclusionAI thanks developers for sharing Ming-3.0
  4. SGLang-Omni supports Ming-3.0 models

    Ant Bailing: Ant Bailing's Ming-3.0 receives day-one support from SGLang-Omni
  5. vLLM-Omni supports deployment of the Ming-3.0 model

    inclusionAI: inclusionAI open-sources five speech models in the Ming-3.0 series, with vLLM-Omni providing Day-0 support
  6. Atomic Chat supports running Ant Bailing's Ming-3.0 speech models locally

    Ant Bailing: Ant Bailing's Ming-3.0 integrates into local AI workflows
  7. Ming-3.0-tts tops the TTS leaderboard

    Artificial Analysis: Ming-3.0-tts tops the Artificial Analysis open-source TTS controllable speech arena and ranks first in pronunciation robustness
  8. Ant's inclusionAI open-sources the Ming-3.0-tts speech synthesis model

    Ant inclusionAI: New GitHub repository: Ant inclusionAI open-sources the Ming-3.0-tts speech synthesis model

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.