Ant Group's Bailing open-sources 5 speech models: 500 million parameter TTS takes first place among open-source in same-timbre blind listening test.

Beating AI News Flash: Ant Ling open-sources the Ming-3.0 speech series, covering three major directions: 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 complete speech input, content analysis, and voice replies.

The most eye-catching is Ming-3.0-tts-0.5B, which has only 500 million parameters. Users can, like a director, use natural language to ask the AI to speak while suppressing anger, emphasize specific words, or adjust the rhythm of lines. It also supports voice cloning, text-designed timbres, and inserting 14 types of sound tags such as light laughter and sighs at specified positions. In Artificial Analysis's same-timbre blind listening test, this model ranked first among open-source models with 1069 Elo.

Ming-3.0-audio-flash, responsible for understanding recordings, is much larger, with total parameters reaching 126 billion and about 6.5 billion activated each time. It can directly receive meeting recordings and answer questions around the discussion content, organizing consensus, disagreements, and to-do items without first separately generating a verbatim transcript.

Another model, Ming-3.0-asr-tiny, specializes in speech recognition and supports inputting personal names, place names, industry terminology, and conversational context to assist in recognizing dialects and uncommon words. In the four types of speech recognition tests published by Ant, its average error rate was 4.41%.