On March 6th, Mobvoi, in collaboration with leading academic institutions including the Hong Kong University of Science and Technology, Shanghai Jiao Tong University, Nanyang Technological University, and Northwestern Polytechnical University, open-sourced the new generation speech synthesis model, Spark-TTS, and launched its commercial high-quality TTS engine—TicVoice7.0.
As Mobvoi's seventh-generation TTS engine, TicVoice7.0 represents a significant breakthrough in speech synthesis, ushering in a new paradigm for speech generation. Its core advantage lies in its innovative speech coding method and modeling structure. The engine employs BiCodec encoding technology, encoding speech into two complementary parts: Global Tokens with a fixed sequence length, and low-bitrate Semantic Tokens. Global Tokens model time-invariant global features such as timbre, ensuring global controllability of speech generation; Semantic Tokens, using features extracted by wav2vec2.0 as input, encode information closely related to the text, ensuring strong semantic relevance. This design not only addresses issues in traditional speech coding but also achieves high unification of speech token modeling and text token modeling, making speech generation more efficient and controllable.
Based on this innovation, TicVoice7.0 demonstrates exceptional voice cloning capabilities and emotional expressiveness. It can accurately capture voice characteristics within 3 seconds, enabling the AI to not only "speak like a human" but also mimic subtle emotional expressions such as sighs and pauses. Compared to the previous generation of large speech models, TicVoice7.0 shows significant improvements in voice similarity, emotional expression, and stability. The internationally recognized MOS score has increased from 3.9 to 4.2, exhibiting stronger emotional expressiveness and a more natural, pleasant, and stable listening experience.
Furthermore, TicVoice7.0 excels in personalized customization. Users can precisely shape unique voice styles by adjusting attributes such as gender, speed, and fundamental frequency. For the "Supreme Pro - Premium Voice Talent" customization, users only need to provide 20-200 sentences of audio to obtain a professional broadcasting-level dubbing experience. The international MOS score has improved from 4.3 to 4.7, reaching broadcast-quality standards and providing a professional-grade speech synthesis solution for film, games, and other applications.
Currently, Mobvoi has integrated TicVoice7.0 into its AI dubbing product, "Moyin Workshop," to provide users with better service and experience. This engine not only performs exceptionally well in applications such as customer service, audiobooks, emotional live streaming, and film narration, but also, through its open-source ecosystem and deep collaboration between industry, academia, and research, injects new impetus into industry development.