- Tencent reports approximately three seconds of translation lag on its RealSI evaluation.
Streaming translation
Tencent Hunyuan released Hy-SimulMT on October 26 for bidirectional simultaneous translation between Chinese and English. Its project documentation says the system receives audio continuously and translates while the speaker is talking, with real-time voice cloning intended to preserve the speaker’s vocal characteristics.
The repository includes the SimulInfer inference service and SimulTrain training tools. Speech output and voice cloning require a configured text-to-speech service, according to the documentation. Tencent also offers a way to try the model in its Hy Translate app.
Vendor-reported evaluation
Tencent reports approximately three seconds of StreamLAAL lag in its RealSI long-form audio evaluation. The documentation explicitly distinguishes this translation-lag metric from time to first speech output. It says its GEMBA translation-quality score exceeded the open-source models and four of five commercial APIs in that comparison.
These are Tencent’s own evaluation results. The release makes the model and supporting tools available for developers working on live bilingual communication, including meetings, speeches and interviews.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
