Tencent Hunyuan has released Hy ASR 3.0 preview, and it's already live in Tencent's Yuanbao app and on Tencent Cloud. According to the company's own benchmarks, the model posts word error rates of 3.34% for Mandarin, 2.62% for English, and 3.12% for Cantonese.
The upgrade connects Tencent's Hy3 large language model to the speech recognition engine. The system now evaluates an entire sentence when deciding between homophones, rather than relying on individual word pronunciation alone. That means when a sentence offers multiple same-sounding candidates, the model uses surrounding context to pick the more appropriate characters.
Hy ASR 3.0 also gets better at dialects, technical terms, mixed Chinese-English speech, background noise, and whispers. Training data includes tens of millions of hours of audio across 10 major dialect regions and more than 20 sub-regions.