Chinese engineers developing chips for artificial intelligence that can be used in "advanced weapons systems" have gained access to cutting-edge UK technology, the Guardian can reveal. Described by analysts as "China's premier AI chip designers", Moore Threads and Biren Technology are subject to US export restrictions over their development of chips that "can be used to provide artificial intelligence capabilities to further development of weapons of mass destruction, advanced weapons systems and hi-tech surveillance applications that create national security concerns". However, prior to the US blacklisting in 2023, the two companies secured extensive licences with the UK-based Imagination Technologies, which is among a handful of firms worldwide that design an advanced type of microchip crucial for AI systems, and is regarded as a jewel of the UK's technology industry. A spokesperson for Imagination said: "At no stage has Imagination (or its owners) considered or implemented transactions with third parties with the aim of enabling China or any other nation state to use or direct Imagination technology for state or military end uses." While Imagination's representatives confirmed the existence of the licences with Moore Threads and Biren Technology, they denied claims that the company, under the ownership of a private equity fund backed with Chinese state money, sought to deliberately transfer its state-of-the-art secrets to China. Two former senior Imagination insiders claim that "knowledge transfer programmes" accompanying the licences were so comprehensive that they risked the Chinese companies learning how to replicate Imagination's expertise.

In this paper, we present MooER, a LLM-based large-scale automatic speech recognition (ASR) / automatic speech translation (AST) model of Moore Threads. A 5000h pseudo labeled dataset containing open source and self collected speech data is used for training. We achieve performance comparable to other open source models trained with up to hundreds of thousands of hours of labeled speech data. Meanwhile, experiments conducted on Covost2 Zh2en testset suggest that our model outperforms other open source Speech LLMs. A BLEU score of 25.2 can be obtained. The main contributions of this paper are summarized as follows. First, this paper presents a training strategy for encoders and LLMs on speech related tasks (including ASR and AST) using a small size of pseudo labeled data without any extra manual annotation and selection. Second, we release our ASR and AST models and plan to open-source our training code and strategy in the near future. Moreover, a model trained on 8wh scale training data is planned to be released later on.