Five speech models released at once
inclusionAI (Ant Ling), the AI division of Chinese financial giant Ant Group, released the weights (trained parameters) of "Ming-3.0," five models that handle speech, on Oct. 19. @aitrackerbot, an X account that tracks the appearance of models, reported the same day that the weights of the speech synthesis model "Ming-3.0-tts-tiny" had been released. All five are under the MIT license, which allows commercial use, and are available from Hugging Face and ModelScope.
Of the five, two are speech synthesis models that read text aloud, two are models that understand the content of speech, and one is a speech recognition model that converts speech to text. Ming-3.0-tts-tiny is a "mixture of experts (MoE)" model that uses only some of its many expert components, with 7.9 billion total parameters, of which only 1.3 billion are used in processing. The largest, the speech understanding model "Ming-3.0-audio-flash," has 126 billion total parameters and handles summarizing long recordings and spoken question answering.
| Model | Use | Total parameters (billion) | Parameters used in processing (billion) |
|---|---|---|---|
| Ming-3.0-tts-0.5B | Speech synthesis | 0.5 | 0.5 |
| Ming-3.0-tts-tiny | Speech synthesis | 7.9 | 1.3 |
| Ming-3.0-audio-flash | Speech understanding | 126 | 6.5 |
| Ming-3.0-audio-tiny | Speech understanding | 9.3 | 2.7 |
| Ming-3.0-asr-tiny | Speech recognition | 8.6 | 2 |
500-million-parameter synthesis model tops open models
U.S.-based Artificial Analysis, which compares AI models under the same conditions, announced on Oct. 19 (Oct. 20 Japan time) that the 500-million-parameter "Ming-3.0-tts-0.5B" ranked first among models with published weights in an evaluation of speech synthesis in which the model is instructed on tone of voice. In the pronunciation accuracy evaluation, it scored 93.2%, tying for first with Fish Audio's "Drama 3 Preview" and outperforming U.S.-based ElevenLabs' "Eleven v4." In an evaluation comparing voices provided by voice companies, it ranked third.
The 500-million-parameter model reads seven languages plus Sichuanese and Cantonese, and 14 kinds of vocal expressions, such as laughter and sighs, can be specified within the text. It can also clone a voice from a short audio clip and create a voice described in text. However, it was this 0.5B version that ranked first in the evaluation, and no ranking was given for the tts-tiny version.
Small and easy to run, aimed at enterprise integration
Speech synthesis is an essential component of AI voice assistants and automated phone answering. Models that use fewer parameters in processing are easier to run even on inexpensive computing resources. Ant made the models usable with the inference software SGLang and vLLM at the same time as the release.
U.S. companies also made moves in voice AI. On the following day, Oct. 20, ElevenLabs added a new model for voice agents, "Eleven Realtime v1," to its official documentation, and France's Mistral AI announced its voice model "Voxtral 3" the same day.
