Skip to main content

Qwen3.8-LiveTranslate Cuts Translation Lag to 2.3 Seconds Per Character

Qwen3.8-LiveTranslate: Real-Time Translation That Finally Feels Live

Ever tried following a live speech through an interpreter and felt like you're always a beat behind? Alibaba's Qwen team just made that gap a whole lot smaller. Their new model, Qwen3.8-LiveTranslate, slashes the average latency per character (LAAL) from 2.8 seconds to 2.3 seconds. Half a second might not sound like much, but in simultaneous interpretation, it's the difference between natural flow and frustrating lag.

More Than Just Speed

The update isn't only about shaving off milliseconds. The model, which already supports 60 languages, now adds three practical features that make real-time translation actually usable.

First, it handles real-time speaker separation. No more guessing who said what—each sentence gets assigned to the right person, and voice replication is more stable. That means no more crosstalk or confusing switches between speakers.

Second, it displays source and translated text simultaneously. You can read along while you listen, which is a lifesaver for anyone trying to catch names, numbers, or technical terms on the fly.

Third, it improves long-context disambiguation. The model uses previous context to understand what's being said now, which dramatically reduces errors on proper nouns and jargon—the stuff that usually trips up live translation.

Image

Under the Hood: Thinker and Talker

The architecture is built on a Hybrid MoE design with two modules: a Thinker and a Talker. The Thinker takes video, audio, original text, and translation, lines them up in a single causal sequence, and generates results end-to-end. In plain English: it understands and translates in one go. The Talker then takes that translated text and the original audio to synthesize speech that keeps the speaker's own voice. So you're not hearing a generic robot voice—you're hearing the actual speaker, just in another language.

Image

Why It Matters

Live interpretation has always been a trade-off between speed and accuracy. You either get a quick, rough translation or a precise one that arrives too late. Qwen3.8-LiveTranslate tries to close that gap. With a 0.5-second improvement per character, the listening experience becomes noticeably smoother. And features like speaker separation and simultaneous text display address real pain points that anyone who's used live translation will recognize.

For developers, this opens up new possibilities for building real-time translation into apps, conferences, and streaming platforms. For everyday users, it means one step closer to seamless communication across languages.

Key Points

  • Latency reduced from 2.8 to 2.3 seconds per character—a 0.5-second improvement.
  • Supports 60 languages with new features: real-time speaker separation, simultaneous source/translated text display, and improved long-context disambiguation.
  • Hybrid MoE architecture with Thinker (understanding + translation) and Talker (voice synthesis) modules.
  • Voice preservation: output speech retains the original speaker's voice.
  • Practical impact: smoother live interpretation, fewer errors on names and terminology, and better usability for developers and end-users.