Qwen3.8-LiveTranslate, the new real-time speech translation model from Alibaba Qwen, reimagines real-time simultaneous interpretation with a unified Interleave architecture. This enhances accuracy, fluency, and conciseness, reducing the average latency per character (LAAL) from 2.8 seconds to 2.3 seconds—this 0.5-second reduction directly determines whether the listening experience feels natural for a field where "a second's delay makes you fall behind."
Building on its support for 60 languages, the new model adds three key pieces that make real-time interpretation truly practical. First, it enables real-time speaker separation, assigning each sentence to the correct person, with more stable voice replication, eliminating issues like confusion between speakers or audio cross-talk. Second, it displays source and translated text simultaneously, allowing listeners to follow along as they listen. Third, it improves long-context disambiguation, enabling the model to understand current content based on previous context, significantly improving accuracy for names and terminology, which are most prone to errors.
The foundation is a Thinker–Talker dual-module design based on Hybrid MoE, using Interleave to connect streaming understanding, text output, and speech generation into one pipeline. The Thinker organizes video, audio, original text, and translation into a single causal sequence, arranging them in time order and generating end-to-end results, completing both "understanding" and "translation" within the same sequence. The Talker then takes the translated text and source audio to synthesize the translation into speech that retains the original speaker's voice—that is, the voice of the speaker remains the same.
