Qwen expands LiveTranslate to 60 languages
Alibaba's Qwen team has released Qwen3.5-LiveTranslate-Flash, a simultaneous interpretation model that takes in 60 languages, uses visual context and clones the speaker's voice, with latency cut to 2.8 seconds per token.
The Qwen team announced Qwen3.5-LiveTranslate-Flash, a real-time interpretation model built on Qwen3.5-Omni. It translates speech and also uses video input to improve accuracy.
Compared with Qwen3-LiveTranslate, the new model accepts speech input and writes text output in 60 languages, up from 18. It speaks 29 output languages, up from 10. The additions include lower-resource languages such as Lingala, Maori, Uyghur and Odia.
Speed is the main improvement. Qwen credits a streaming method it calls "Readable Unit," which tags the output to control how finely speech is synthesized. The team says this cuts first-token latency by 3.45 seconds and per-token latency by 1.88 seconds compared with the previous Flash model. Average speech-to-speech per-token latency is now 2.8 seconds, and Qwen says translation quality is "virtually" unchanged.
The model uses the Thinker-Talker design from Qwen3.5-Omni. The Thinker takes in mixed audio and visual input and writes the translated text. The Talker then turns that text, along with the source audio, into speech in the speaker's own voice. Voice cloning is new in this release and has three modes: pre-registered, clone-once and real-time. The model also supports up to 1,000 "hotwords," a list of names and industry terms that users can update while it runs.
On the FLEURS and CoVoST2 speech translation benchmarks, Qwen says the model beats "mainstream commercial large speech models." The post does not give scores or name those competitors. It also does not give pricing or say whether the model weights will be released. The post shows a DashScope demo, which suggests the model is offered through an API.