HONOR has officially announced that its Magic V5 will have the on-device multilingual call translation system—a leap forward in both speed and accuracy. This innovation marks a significant milestone in multilingual speech recognition and translation, addressing long-standing technical challenges such as low-latency streaming recognition and the efficient deployment of large-scale models on mobile hardware.

Privacy Meets Performance

Traditional translation tools often rely on the cloud, raising serious privacy concerns, particularly for sensitive conversations like phone calls. On-device solutions exist, but they often compromise speed, accuracy, or efficiency due to limitations in mobile hardware. HONOR’s new technology changes the equation. By delivering cloud-level translation performance entirely on-device, it eliminates privacy risks while maintaining top-tier speed and accuracy.

Smarter Storage and Real-Time Translation

One of the standout achievements is the memory efficiency. HONOR has slashed the footprint from 3–4GB to just 800MB, saving around 75% of storage space. Six language packs—Chinese, English, German, French, Spanish, and Italian—are embedded directly, avoiding the need for bulky downloads and saving roughly 2.78GB of storage.

The Magic V5 also introduces “speak-as-you-go” real-time translation, moving beyond traditional methods that wait for full sentences. This improvement boosts inference speed by 38% and translation accuracy by 16%, making conversations more natural and seamless.

Backed by INERSPEECH 2025 Research

HONOR’s announcement is underpinned by peer-reviewed research presented at INTERSPEECH 2025.

• The first paper, MFLA: Monotonic Finite Look-ahead Attention for Streaming Speech Recognition, details how HONOR adapted the Wait-k strategy (common in text translation) to speech recognition by combining it with a Continuous Integrate-and-Fire (CIF)-based predictor, drastically cutting latency while preserving accuracy.

• The second paper, Novel Parasitic Dual-Scale Modelling for Efficient and Accurate Multilingual Speech Translation, introduces a parasitic dual-scale speculative sampling approach (developed with Shanghai Jiao Tong University) that accelerates inference by 38% on-device without compromising quality.

The Future of On-Device AI

With the Magic V5, HONOR signals its commitment to private, real-time, AI-driven communication. By translating faster, lighter, and more secure—without the cloud—this breakthrough sets the stage for a new era of intelligent human-device interaction.