Qwen Live Translate: Real-Time Multimodal Speech Translation

Model Overview

Qwen Live Translate (Qwen3.5-LiveTranslate-Flash) is Alibaba's real-time multimodal simultaneous speech-to-speech and speech-to-text translation model family built on the Qwen-Omni architecture by the Qwen Team (Alibaba Tongyi Lab).

It fuses audio, text, and visual input (such as speaker lip movements and facial context) to provide low-latency simultaneous interpretation, cross-lingual voice cloning, and domain-specific terminology control for live streaming and international conferences.

---

Key Features

• Multimodal Audio-Visual Understanding: Integrates audio, text, and visual cues (lip movements, on-screen text) to resolve translation ambiguities in noisy environments.
• Ultra-Low Latency Streaming: Utilizes "Readable Unit" chunk-based streaming to achieve average end-to-end speech-to-speech latency as low as 2.8 seconds.
• Multilingual Support: Supports 60 languages for audio input and text translation, and 29 languages for speech audio synthesis.
• Cross-Lingual Zero-Shot Voice Cloning: Automatically replicates the speaker's vocal characteristics (timbre and voice identity) in real-time.
• Hotword Customization: Configures custom hotwords and domain terms to guarantee high accuracy in technical contexts.

---

Verified Project Links

• Project Portal: https://qwenlm.github.io/
• arXiv Paper: https://arxiv.org/abs/2604.15804
• GitHub Repository: https://github.com/QwenLM
• Hugging Face Hub: https://huggingface.co/Qwen

---

Performance & Benchmarks

• FLEURS & CoVoST 2: Achieves state-of-the-art BLEU/COMET scores across major speech translation directions.
• End-to-End Latency: 2.8 seconds average end-to-end latency.