Tencent Hunyuan releases Hy ASR 3.0: lowest aggregate error on public Mandarin, English and Cantonese benchmarks among evaluated systems
Synopsis
Tencent Hunyuan released Hy ASR 3.0, a large-language-model-based speech recognition system covering Mandarin Chinese, English, Cantonese and 12 other Chinese dialect groups, with external context supplied as hotword lists, question-answer dialogues or summaries. On public Mandarin, English and Cantonese benchmarks it reaches aggregate error rates of 4.64%, 5.22% and 5.06%, each below the evaluated non-Hy systems; the Cantonese-only aggregate is about 11.0% lower than Qwen3 ASR Flash. The system also uses information inside the same utterance to disambiguate homophonic or phonetically similar keywords without a separate context input.
Interpretation
Hy ASR 3.0 brings the aggregate error rate on all three public benchmark groups, Mandarin, English and Cantonese, to the lowest among evaluated systems, at 4.64%, 5.22% and 5.06%. The strongest non-Hy comparators were previously FunASR Flash at 4.77% for Mandarin, StepAudio 3 ASR Max at 5.38% for English and Qwen3 ASR Flash at 5.68% for Cantonese. Against those three strongest non-Hy comparators, the relative error reductions are about 2.7%, 3.0% and 11.0% for Mandarin, English and Cantonese, with the Cantonese aggregate excluding KeSpeech.
The system combines an audio encoder, an audio adapter and an LLM decoder with an audio-conditioned hotword retrieval mechanism. Hotword retrieval selects terms relevant to the current speech from a candidate list, and these are encoded with any summary or dialogue context as a textual prefix placed before the audio embeddings. The LLM decoder generates the transcription autoregressively conditioned on that context prefix, the audio embeddings and previously generated tokens, while intra-utterance context is carried by the audio and transcription sequence rather than supplied as a separate external input.
Training runs through encoder pre-training, large-scale ASR pre-training and post-training, where three domain experts are merged into one model by multi-teacher on-policy distillation. The three experts are trained separately with reinforcement learning for general, dialectal and context-biasing recognition, which reduces cross-domain interference. For each utterance the student first generates an on-policy hypothesis, which is routed to the matching domain expert for token-level supervision; a final model-behavior reinforcement learning stage targets hard cases from real usage, especially keywords whose recognition depends on surrounding context.
Perspective
Evaluation covers public Mandarin, English, Cantonese and dialect benchmarks plus internal Mandarin, English, Cantonese and code-switching test sets built from real business data; the dialect portion adds human evaluation over 13 dialect categories and 91 subregional varieties. Robustness conclusions are limited to three internal settings: long-form audio, a combined child-and-older-adult test set, and a combined low-volume and noisy test set.
The Mandarin aggregate advantage does not hold on every subset, and FunASR Flash remains stronger on AISHELL-1 and WenetSpeech test-meeting; in English, StepAudio 3 ASR Max remains stronger on both LibriSpeech subsets and the accented VoxPopuli-en subset, where Hy ASR 3.0 also regresses from 14.75% to 15.27% relative to Preview. In the dialect human evaluation, Hy ASR 3.0 trails FunASR Flash by 1.0 percentage point under the lenient criterion, and FunASR Flash leads on Wu under all three metrics; the combined child-and-older-adult result does not establish separate improvements for each age group.