HONOR has announced a major AI breakthrough with the debut of the industry’s first on-device large speech model on the international versions of the HONOR Magic V5, set to launch later this week. This innovation addresses longstanding challenges in multilingual speech recognition and translation, marking an important step in bringing cloud-level AI performance directly to smartphones.
Solving the Cloud vs. Device Trade-Off
Mainstream translation and voice recognition technologies today rely heavily on cloud infrastructure, raising privacy concerns—especially for sensitive conversations such as phone calls. While some on-device solutions exist, they often face compromises in speed, accuracy, and memory efficiency.
HONOR’s new on-device large speech model aims to overcome these hurdles. By deploying large-scale models directly on smartphones, it provides a cloud-comparable experience while ensuring greater data privacy.
Breakthroughs in Technology
According to HONOR, the solution integrates multiple advancements:
- Smaller storage footprint: Six language packs (Chinese, English, German, French, Spanish, Italian) have been compressed from 3–4GB to just 800MB, saving users roughly 2.78GB of storage.
- Speak-as-you-go translation: Instead of waiting for full sentences, the system can translate in real time as users speak, making conversations smoother and faster.
- Performance boost: Tests show a 38% increase in inference speed and a 16% boost in translation accuracy compared with conventional systems.
Recognized by INTERSPEECH 2025
Two academic papers highlighting these advancements have been accepted at INTERSPEECH 2025, the world’s leading conference on spoken language processing.
- “MFLA: Monotonic Finite Look-ahead Attention for Streaming Speech Recognition” – Tackles low-latency, high-accuracy speech recognition on devices using an innovative Continuous Integrate-and-Fire (CIF)-based predictor combined with the Wait-k strategy.
- “Novel Parasitic Dual-Scale Modeling for Efficient and Accurate Multilingual Speech Translation” – Developed with Shanghai Jiao Tong University, this research introduces a dual-scale speculative sampling acceleration method that improves inference speed by 38% while keeping accuracy intact.
Towards Smarter, Private AI on Devices
HONOR emphasized that this development is part of its broader commitment to on-device AI, which places user privacy and real-time performance at the forefront. By embedding advanced speech models directly into devices, the Magic V5 positions itself as a milestone for intelligent, secure, and seamless human-device interactions.
The HONOR Magic V5’s global debut later this week is set to showcase not just a new foldable smartphone, but also one of the most ambitious integrations of AI into consumer devices to date.