Dubai - Emirates247: HONOR has launched a large speech language model that runs directly on-device as part of the international release of its new Magic V5 smartphone. The technology focuses on multilingual speech recognition and real-time translation without relying on cloud infrastructure or requiring an internet connection. Features and EfficiencyAccording to the company, the model enables speech recognition and instant translation with low latency, while reducing memory consumption to around 800 MB, compared to the previous requirement of 3–4 GB. The first paper, titled “MFLA: Monotonic Finite Look-ahead Attention for Streaming Speech Recognition”, introduces an approach combining Continuous Integrate-and-Fire (CIF) prediction with a Wait-k strategy to enable low-latency and high-accuracy speech recognition on devices. The second, in collaboration with Shanghai Jiao Tong University, presents a dual-scale “speculative sampling” model designed to accelerate multilingual speech translation while maintaining efficiency on resource-limited devices.