HONOR Showcases Breakthrough On-Device AI Speech Technologies on the Latest Flagship, Magic V5

(MENAFN- mslgroup)
Global leading AI device ecosystem company HONOR announced the debut of the industry's first on-device large speech model on the international versions of the HONOR Magic V5. This achievement marks the successful resolution of key technical challenges in multilingual speech recognition and translation on devices, including major breakthroughs in low-latency streaming speech recognition and the efficient deployment of large-scale models.

Supporting this technological advancement, two related research papers have been recognized at INTERSPEECH 2025, the world’s largest and most comprehensive conference on the science and technology of spoken language processing.

Addressing the Privacy and Performance Dilemma
Current mainstream translation solutions are heavily reliant on cloud infrastructure, raising significant privacy concerns, especially for sensitive conversations like phone calls. While some existing on-device solutions available on the market attempt to address this, they often compromise significantly on performance, including speed, accuracy, and memory footprint, due to the inherent limitations of mobile devices. HONOR's new technology decisively overcomes these limitations, delivering a cloud-comparable experience directly on-device, thereby ensuring both robust privacy and superior performance.

Unlocking Unprecedented On-device Communication Benefits
HONOR's innovative solutions bring a host of core consumer benefits. It achieves drastic memory efficiency, reducing the memory footprint from 3-4GB to a mere 800MB, saving an impressive 75% of storage. This includes embedding six language packs (Arabic, Chinese, English, German, French, Spanish, and Italian), eliminating the need for six separate 500MB downloads and saving approximately 2.78GB of storage. The technology enables "speak-as-you-go" real-time translation, a significant advancement over traditional methods that require waiting for a full sentence to be completed, resulting in a 38% increase in inference speed and a 16% increase in translation accuracy.

INTERSPEECH 2025 Validates Groundbreaking Research
The first paper, "MFLA: Monotonic Finite Look-ahead Attention for Streaming Speech Recognition," addresses the critical challenge of achieving low-latency and high-accuracy streaming speech recognition on devices. HONOR's innovative integration of a Continuous Integrate-and-Fire (CIF)-based predictor with the Wait-k strategy is a key highlight. While traditional Wait-k strategies perform well in discrete token tasks like machine translation, their direct application to Automatic Speech Recognition (ASR) is limited by the continuous nature of speech, leading to high computational costs. HONOR introduced a predictor based on the CIF mechanism. This predictor explicitly maps continuous acoustic features to the discrete boundary decisions required by the Wait-k strategy, successfully adapting this low-latency approach from the text domain to the speech domain.

The second paper, "Novel Parasitic Dual-Scale Modeling for Efficient and Accurate Multilingual Speech Translation," overcomes the limitations of real-time inference for large speech models on resource-constrained devices. This paper introduces a parasitic dual-scale speculative sampling acceleration strategy, developed in collaboration with Shanghai Jiao Tong University, which can be deployed on edge devices and achieves a 38% increase in inference speed without compromising model performance.

HONOR remains steadfast in its commitment to pushing the boundaries of on-device AI. This innovative technology paves the way for more intelligent, private, and seamless human-device interactions in the future.

MENAFN31082025004993017291ID1109997714