Honor Magic V5 debuts with the first Large Speech Model On-Device for faster, safe and precise translations

Honor is preparing for a launch of great international importance in the field of artificial intelligence: the new Magic V5 Foldable coming in a few days, introduces the first large speech model on-device Never seen on a smartphone.

A novelty that not only marks a step forward in artificial intelligence applied to language, but deals with two central issues in the world of automatic translation in a concrete way, namely the privacy and real -time performance.

So far, most of the translation and vocal recognition solutions has reached the cloud, with all the limits and critical issues that this entails both from the point of view of the processing speed (any request is forwarded to the cloud which returns it to the device) and of privacy of its sensitive data.

With Magic V5, Honor tries to raise the bar in the sector by bringing ability on the device that, until yesterday, seemed unthinkable for today’s smartphones.

Follow Honor Italia on Telegram, Receive news and offers first

Honor’s on-device vocal model marks a turning point, that’s why

The cloud-based solutions have so far dominated the market thanks to their calculation power and the possibility of constantly updating linguistic models.

However, they involve Two significant problems. The first is privacy: send potentially sensitive conversations to external servers raises doubts about the protection of personal data. The second is linked to dependence on network connection: if the signal is weak or absent, the translation experience is compromised or simply impossible.

The so-called on-device solutions already existing, however, while offering greater safety, have always shown limits in terms of processing speed, accuracy of recognition and memory consumption.

It is precisely here that Honor comes into play with the debut of the Large Speech Model On-Devicedeveloped to guarantee an experience comparable to that of the cloud, but entirely local on the device And therefore more reliable and safe.

Fast and more precise translations with less memory consumption

According to data shared by the company, the new system allows you to drastically reduce the weight of linguistic packages. Translating six main languages, Chinese, English, German, French, Spanish and Italian, traditionally required 3 to 4 GB of overall memory on the smartphone.

With the new technology, the same function is managed in just 800 MBfreeing over 2.78 GB of space on the device.

According to the company, this system integrates the “Speak-Aas-You-Go“: That is, the translation takes place as we speak, without having to wait for the completion of the sentence.

In concrete terms, Honor declares a 38% increase in processing speed It is an improvement of 16% in precision of translations compared to previous methods.

Scientific recognition: two studies rewarded in Interspeech 2025

This innovation is not limited to marketing, but also finds confirmation in the scientific community.

Indeed, Two research articles which describe in detail the technologies behind the new system have been selected and recognized in Interspeech 2025the most important world conference on spoken language science and technology.

The first paper, entitled “MFLA: Monotonic finished Look-ahead attention for Streaming Speech Recognition”, Faced the latency problem in vocal recognition in real time.

To solve it, Honor has integrated a predictor based on Continuous integrated-and-fire (CIF) with the Wait-K strategy, an approach originally developed for the text and successfully adapted to spoken language.

The second study, “Novel Parasitic Dual-Scale Modeling for Efficient and accurate multilingual Speech Translation“, It is the result of a collaboration with the Shanghai Jiao Tong University.

Here an innovative acceleration strategy is presented based on Parasitic Dual-Scale Speculative Sampling, designed for optimize inference in real time also on mobile devices with limited resources. The result is an increase of 38% in the processing speed without sacrificing the quality of the translation.

In addition to the technical details, what makes this debut particularly interesting is obviously the benefits in daily use. With Magic V5, Honor aims to offer a tool that breaks down linguistic barriers in real time, minimizing compromises.

On-device to the Honor strategy pillar

The introduction of the Large Speech Model on the Magic V5 is part of a wider strategy that has seen Honor for some time engaged in investing onTo the on-device as a technological differentiation compared to the competition.

The debut of the first Large Speech Model On-Device on the Magic V5 marks an important moment for Honor and, more generally, for The evolution of mobile artificial intelligence.

Such innovation represents a milestone in the way AI is conceived and implemented on mobile devices and we are sure that this Honor step is the turning point that will be able to start an increasingly widespread adoption of this technology throughout the sector.