Qwen3.8-LiveTranslate - Alibaba Tongyi's Real-Time Simultaneous Interpretation Model

What is Qwen3.8-LiveTranslate

Qwen3.8-LiveTranslate is a real-time simultaneous interpretation large model launched by Alibaba Tongyi Qianwen. It is rebuilt with the Interleave architecture, reducing the average per-character latency from 2.8 seconds to 2.3 seconds, and supports 60 languages. The model has three major capabilities: real-time speaker diarization and voice cloning, simultaneous output of source text and translation in the same frame, and long-context disambiguation. It can recognize video/audio input and output bilingual text and translated speech that preserves the original voice end-to-end. It is suitable for scenarios such as multinational conferences, cross-border livestreaming, and video localization, and its API is already available through the Qianwen AI platform.

Main Features of Qwen3.8-LiveTranslate

• Real-Time Simultaneous Interpretation: Audio-text single-stream processing based on the Interleave architecture, with average per-character latency of only 2.3 seconds, translating while listening.
• Speaker Diarization: Automatically distinguishes each speaker when multiple people speak in turn, and separately clones their voices to generate translated speech.
• Bilingual Same-Frame Output: Source text and translation are presented synchronously on the same screen, making it easy to understand instantly and verify against the original.
• Long-Context Disambiguation: Understands the current sentence in combination with preceding context, making translations of names, terminology, and references more accurate and consistent.
• Multilingual Translation: Supports speech recognition input in 60 languages and speech output in 29 languages.
• Visual Information Assistance: Supports video/image input, using visual content to assist translation disambiguation.

Technical Principles of Qwen3.8-LiveTranslate

• Interleave Single-Stream Architecture: Qwen3.8-LiveTranslate reconstructs simultaneous interpretation into a single stream in which audio and text are interleaved. Heard audio and already-produced translations are both cached and reused, avoiding processing each sentence from scratch, improving translation quality while compressing the average per-character latency from 2.8 seconds to 2.3 seconds.
• Thinker–Talker Dual-Module Design: The model adopts a Thinker–Talker structure based on Hybrid MoE. Thinker arranges video, audio, source text, and translation into the same causal sequence, alternating in temporal order to complete understanding and translation end-to-end; Talker combines the translation with the source audio to synthesize translated speech that preserves the original speaker's voice.
• Real-Time Speaker Diarization: In scenarios where multiple people speak in turn, the model first completes Diarization before translation, assigning each sentence to a specific speaker, and uses this to guide Talker in stably cloning that speaker's voice, ensuring that the translated speech "sounds like who said it."
• Long-Context Disambiguation: The model incorporates historical dialogue into context modeling, associating preceding information across turns, avoiding ambiguity in proper nouns, names, and references caused by judging from a single sentence alone, and ensuring consistency of terminology and expression throughout long multi-person conversations.
• Bilingual Synchronous Output Mechanism: In the architectural design, source text and translation are aligned and generated in the same interleaved sequence, naturally supporting same-frame output, so downstream applications such as bilingual subtitles, content organization, and retrieval require no additional alignment processing.

How to Use Qwen3.8-LiveTranslate

• Open the URL https://omni.qwen.ai/live-translate。
• Allow microphone permission, and select the source language and target language.
• Start speaking directly, and you can see real-time bilingual subtitles and hear the translated speech read aloud.

• Log in to the Qianwen AI platform and obtain an API Key.
• Call the https://www.qianwenai.com/models/qwen3.8-livetranslate-flash-realtime model's real-time interface and pass in the audio stream.
• Simply receive the returned bilingual text and translated speech.

Core Advantages of Qwen3.8-LiveTranslate

• Low Latency: The Interleave architecture reduces per-character latency from 2.8 seconds to 2.3 seconds. Already-heard audio and already-produced translations can be cached and reused, making translation faster and smoother.
• Speaker Diarization + Voice Cloning: When multiple people speak in turn, speakers are precisely distinguished, and the translated speech clones each person's voice respectively, making the attribution of communicated information clear.
• Bilingual Output in the Same Frame: The source text and translation are presented synchronously, balancing instant comprehension with source-text verification, naturally suited to scenarios such as subtitles, content organization, and retrieval.
• Long-Context Disambiguation: It associates historical context across turns, resolving ambiguity that cannot be judged from a single sentence, and translating names and terminology more accurately with consistency throughout.
• Broad Coverage of 60 Languages: It supports recognition of 60 languages and speech output in 29 languages, meeting the needs of multilingual real-time communication.

Comparison of Qwen3.8-LiveTranslate with Similar Competing Products

Application Scenarios of Qwen3.8-LiveTranslate

• Multinational Conferences: When multiple people speak, it automatically distinguishes speakers and clones their voices for real-time translation, making participants feel as if they are communicating in the same language.
• Cross-Border Livestreaming: It outputs synchronized bilingual subtitles and translated speech for streamers, helping livestream content reach global audiences without barriers.
• Video Localization: Input a video to generate aligned bilingual subtitles and dubbing, greatly simplifying the translation and production process for film, television, and course content.
• International exhibitions/conventions: Provide low-latency simultaneous interpretation for on-site communication between exhibitors and visitors, reducing the cost of professional interpreters.
• Customer service and training for overseas-expanding enterprises: Support real-time voice inquiries from overseas customers and cross-language employee training, improving the efficiency of globalized services.