According to CoinMeta, Tencent's Hy ASR 3.0 has now been launched on Yuanbao and Tencent Cloud. Official tests show that its error rate for Mandarin words is 3.34%, for English words it is 2.62%, and for Cantonese words it is 3.12%. The new model integrates the Hy3 large language model into the speech recognition system, combining whole sentences to identify homophones and improving recognition accuracy. In addition, Hy ASR 3.0 has also enhanced recognition of dialects, specialized terms, mixed Chinese-English speech, noise, and whispers. The training data covers tens of millions of hours of speech, as well as 10 major dialect regions and more than 20 sub-regions.