A new technical report describes Hy ASR 3.0, Tencent researchers' speech-recognition system, and how it uses contextual hints when transcribing audio. Submitted November 11, the report discusses a model already offered as a preview earlier in 2026; the paper should not be mistaken for a new model launch.
The authors describe support for Mandarin, English and 13 Chinese dialect groups, including Cantonese. They also describe using external information such as hotwords, dialogue and summaries to guide recognition. A supplied vocabulary can help with names or specialized terms, but a hint can also introduce bias if it is wrong or does not apply to the recording.
What the evaluation shows
The report presents results across public and internal speech sets, and examines noisy audio, long recordings and speakers of different ages. The researchers report strong results on their selected comparisons. These are authors' results on specified test conditions, not an independent assessment by AiBrainWorX or a promise of accuracy on every accent, microphone or setting.
The full report explains its audio representations, contextual inputs and training methodology. The report does not establish that weights are publicly available for local use. Readers should check the provider's current availability and terms rather than assume the system is open-weight.
Our view: context-aware transcription is valuable when domain terms are known in advance, but an evaluation should also test when context is absent, misleading or sensitive. We have not tested the service or model.
Original reporting; primary sources reviewed November 14, 2026. No independent hands-on test claimed.
