image
By Asia Education Review Team , Tuesday, 24 September 2024

Jingzhunxue Launches World's First Open-Source Speech LLM for Education

  • Jingzhunxue, an education technology company based in Hangzhou, has introduced 'FlowMirror-s(V02)' the world's first open-source end-to-end large language model (LLM) designed for speech detection and interaction in the education sector. This launch represents the company's commitment to fostering knowledge sharing within the industry and encouraging the adoption of AI-driven learning services.

    Unlike conventional Automatic Speech Recognition and Text-to-Speech systems, 'FlowMirror-s(V02)' is developed from the ground up with a self-supervised Chinese speech codec system, initializing its weights from a text LLM. This approach results in a model that is trained end-to-end on speech and dialogue data, enabling low-latency, seamless speech interactions.

    Jingzhunxue has also developed the world's first 'Hyper-Realistic AI One-on-One Tutor' using its FlowMirror-s(V02) model. Designed under the 'AI Native' principle, this AI tutor closely mimics real teachers, offering personalized, systemized instruction that surpasses traditional AI tools and extends across all scenarios. "We focus especially on educational applications. Hopefully, the success of this model will not only drive innovation in the Chinese AI community but also advance the education sector", says Renbin Yang, founder and CEO of Jingzhunxue. 

    Utilizing Alibaba's LLM Tongyi Qianwen (Qwen), FlowMirror emerges as one of the most advanced educational models in China. Built on Qwen's trillion-parameter framework and trained on AliCloud's multi-GPU Bailian platform, FlowMirror provides a variety of specialized features.

    The model incorporates more than 2 billion proprietary tokens via an advanced data pipeline tailored for educational support. It facilitates multimodal interaction by integrating Alibaba's visual model, which enables dynamic problem-solving and tutoring assistance. Additionally, the model can identify over 40 emotional and physical states using voice and visual cues. Furthermore, with 160,000 hours of educational speech training, it aligns its speech patterns closely with those of human teachers.

    Moreover, FlowMirror provides personalized teaching styles influenced by esteemed educators and employs virtual teacher technology to create unique, high-definition, real-time interactions using just one hour of video data. Its knowledge graph also leverages millions of datasets sourced from a comprehensive question bank and learning data pool, enabling it to master a wide range of educational materials.

    FlowMirror-s v0.1 and v0.2 were pre-trained on 20,000 and 50,000 hours of speech data, respectively, confirming the model's end-to-end speech capabilities and scalability. The latest version, "FlowMirror-s(v0.2)," enables seamless speech input-to-output interactions, enhancing its suitability for educational contexts with natural dialogues that mimic real teachers. Jingzhunxue intends to demonstrate the practical applications of this technology in the near future.

    In May, Jingzhunxue raised nearly RMB 200 million ($28 million) in a funding round exclusively supported by Alibaba. The funds will be utilized to develop and promote its AI Native educational devices, including the Bong series of e-learning tablets, which are currently available on e-commerce platforms such as Tmall and JD.com.