Tencent Hunyuan open-sources the Hy-SimulMT simultaneous interpretation model

Article

Tencent Hunyuan open-sources AI Simultaneous interpretation model: listen-and-translate latency of approximately 3 seconds, with speaker voice cloning

Beating AI Flash reports that Tencent Hunyuan has open-sourced the simultaneous interpretation model Hy-SimulMT, supporting real-time two-way Chinese-English translation. Without waiting for the other person to finish a sentence, the model can listen and generate the translation simultaneously.

After integrating an additional TTS model, it can also clone the original speaker's voice and deliver the speech in another language. The official deployment solution is based on Qwen3-TTS.

In the RealSI long-audio test, the official team measured streaming translation latency of approximately 3 seconds, the lowest among the tested systems. Translation quality surpassed that of all tested open-source models, as well as 5 commercial API out of 4. The 3-second figure does not include the full speech synthesis and playback time.

The team also recruited 6 certified professional interpreters to conduct simultaneous tests on 18 recordings of real speeches, interviews, and other content. In this small-scale test, the model's translation quality reached a professional simultaneous interpretation level, while latency was generally lower.

The model has approximately 30000000000 total parameters, with approximately 3000000000 activated for each inference. The weights, inference tools, and training tools are all open-sourced, and the Tencent Hy Translation app also offers a free trial.

Original link https://m.theblockbeats.info/flash/373947

Key points

Hy-SimulMT supports real-time two-way Chinese-English translation.
The model can listen and generate translations simultaneously; the report cites RealSI tests showing streaming translation latency of approximately 3 seconds.
After integrating an additional TTS model, the voice of the original speaker can be cloned and used to speak in another language.
The model has approximately 30000000000 total parameters, with approximately 3000000000。
The weights, inference, and training tools are all open source, and the Tencent Hy translation app offers a free trial.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a model for real-time simultaneous interpretation that begins generating the translation before the other speaker has finished a sentence. Combined with a speech synthesis model, it can also broadcast the translation in a voice close to that of the original speaker.

Why it matters to readers

It combines real-time translation, low-latency processing, and voice cloning, while also making the model weights and tools available.

Beginners can understand it as a speech model that translates while listening; approximately 3 seconds is the streaming translation latency, which is not the same as the time required for complete broadcasting.

Risks and unknowns

The report did not include an official announcement from Tencent or independently verifiable testing materials.
The approximately 3-second latency does not include the time required for complete speech synthesis and broadcasting.
The professional interpreter testing involved only 6 interpreters and 18 audio recordings, so the sample size was limited.
The test results and open content reported still need to be verified against the official repository, licensing terms, and the complete testing methodology.

Tencent Hunyuan open-sources the Hy-SimulMT simultaneous interpretation model

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center