Ant Ling open-sources the Ming-3.0 speech series, including 5 MIT licensed models

Article

Ant Ling open-sources 5 speech models: 500000000 parameters TTS ranks first among open-source models in same-voice blind listening tests

Beating AI Newsflash: Ant Ling has open-sourced the Ming-3.0 speech series, covering major areas: speech synthesis, audio understanding, and speech recognition, with a total of 5 models, including 2 TTS、2 Audio and 1 ASR, all under the MIT license. Developers can combine them to handle voice input, content analysis, and voice responses.

The standout is 500000000 with only parameters.Ming-3.0-tts-0.5B Users can act like directors, using natural language to instruct AI to speak while suppressing anger, emphasize specific words, or adjust the pacing of lines. It also supports voice cloning, text-based voice design, and inserting 14 types of sound tags, such as light laughter and sighs, at designated positions. In Artificial Analysis's same-voice blind listening test, the model ranked first among open-source models with 1069 Elo.

The Ming-3.0-audio-flash responsible for understanding recordings is much larger, with total parameters reaching 126000000000, of which approximately 6500000000 are activated each time. It can directly receive meeting recordings, answer questions about the discussion, and organize consensus, disagreements, and action items, without first separately generating a verbatim transcript.

Another Ming-3.0-asr-tiny specializes in speech recognition and supports inputting names, place names, industry terminology, and prior dialogue context to assist with recognizing dialects and uncommon words. Across the categories of speech-recognition tests published by Ant, its average error rate was 4.41%.

Original link https://m.theblockbeats.info/flash/372755

Key points

Beating AI reports that Ant Ling has open-sourced the Ming-3.0 speech series, comprising 5 models.
The series includes 2 TTS、2 Audio and 1 ASR models, all under the MIT license.
The report states that Ming-3.0-tts-0.5B has 500000000 parameters and supports natural-language control of expression, voice cloning, text-based voice design, and 14 types of sound tags.
The report states that Ming-3.0-tts-0.5B ranked first among open-source models in Artificial Analysis's same-voice blind listening test with 1069 Elo.
Ming-3.0-audio-flash Total parameters are 126000000000, with approximately 6500000000 activated each time; it can directly process meeting recordings, answer questions about their content, and organize consensus, disagreements, and action items.
Ming-3.0-asr-tiny Supports inputting names, place names, industry terminology, and prior dialogue context; across the categories of speech-recognition tests published by Ant, its average error rate was 4.41%。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This set of models respectively handles “turning text into speech,” “understanding recordings,” and “turning speech into text.” Developers can combine these capabilities to build a workflow from voice input and content analysis to voice responses.

Why it matters to readers

The material presents a set of models covering multiple speech stages, under the MIT license, and provides metrics such as parameter scale, blind-listening ranking, and recognition error rate.

Beginners can first select models by task:TTS for speech synthesis,Audio for understanding recordings, and ASR for speech recognition; before actual use, verify the model files, license scope, and test conditions.

Risks and unknowns

The material does not include an official Ant Ling release page or model repository, so it is not possible to directly verify the model weights,MIT license, or scope of availability.
The material does not provide the original Artificial Analysis test page, so it is not possible to verify 1069 Elo, the ranking, or the specific test conditions.
4.41%The average error rate comes from tests published by Ant; the material does not specify the datasets, calculation methodology, or independent replication results for the test categories.
Although the material states there are 5 models in total, it explicitly lists only Ming-3.0-tts-0.5B、Ming-3.0-audio-flash and Ming-3.0-asr-tiny as names.

Ant Ling open-sources the Ming-3.0 speech series, including 5 MIT licensed models

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center