MiniMax's Speech-2 reached the top of ArtificialAnalysis, an authoritative international voice evaluation list, beating OpenAI and ElevenLabs. Speech-2 uses a TTS model with an autoregressive Transformer architecture to achieve zero-sample speech cloning and a new Flow-VAE architecture, improving the quality and similarity of speech generation. Speech-2 also has characteristics such as superpersonation, individuality, and diversity. It supports 32 languages, and has excellent performance in languages such as Chinese and English. MiniMax is exploring implementation solutions for scenarios such as voice assistants, voice chat, and dubbing, and is speeding up the commercialization process.

MiniMax's Speech-2 reached the top of ArtificialAnalysis, an authoritative international voice evaluation list, beating OpenAI and ElevenLabs. Speech-2 uses a TTS model with an autoregressive Transformer architecture to achieve zero-sample speech cloning and a new Flow-VAE architecture, improving the quality and similarity of speech generation. Speech-2 also has characteristics such as superpersonation, individuality, and diversity. It supports 32 languages, and has excellent performance in languages such as Chinese and English. MiniMax is exploring implementation solutions for scenarios such as voice assistants, voice chat, and dubbing, and is speeding up the commercialization process.