We present Hy ASR 3.0, a large language model (LLM)-based automatic speech recognition (ASR) system that delivers strong recognition accuracy, supports flexible contextual conditioning, and maintains robust recognition performance across diverse speech conditions. Hy ASR 3.0 supports Mandarin Chinese, English and 13 Chinese dialect groups, including Cantonese. Among the evaluated systems, it achieves the lowest aggregate error rates on public Mandarin Chinese, English, and Cantonese benchmarks and the lowest error rates on the corresponding internal test sets, alongside first-tier dialect recognition performance. Beyond audio-only transcription, Hy ASR 3.0 supports multiple forms of external context, including hotword lists, question--answer dialogues, and summaries, enabling the use of task-relevant information during recognition. The model also leverages intra-utterance context to disambiguate homophonic and phonetically similar keywords using information from other parts of the same utterance, without requiring a separate contextual input. Evaluations demonstrate strong contextual recognition in both external-context and intra-utterance settings, as well as robust recognition of long-form speech, children's and older adults' speech, low-volume speech, and speech in noisy environments. This report describes the model architecture and training methodology and presents evaluations across recognition accuracy, context utilization, and robustness.
Figures & tables
Figure 1: Overview of Hy ASR 3.0 capabilities and benchmark performance. Comparisons are made against the strongest competitor for each metric, with the corresponding model shown in parentheses. RERR denotes the relative error rate reduction.
Figure 2: Overview of the Hy ASR 3.0 model architecture and its training procedure.
Testset
Qwen3 ASR Flash
FunASR Flash
Doubao ASR 2.0
Seed 2.0 Lite
StepAudio 3 ASR Max
Hy ASR 3.0 Preview
Hy ASR 3.0
Chinese (zh)
5.73
4.77
5.55
5.42
6.98
4.73
4.64
AISHELL1
2.12
1.01
1.90
2.03
1.40
1.21
1.37
AISHELL2 android/ios/mix
2.70/2.59/2.55
2.16/ 1.91 /2.14
2.67/2.55/2.69
2.68/2.75/2.78
2.30/2.20/2.35
2.02 /2.03/ 2.10
2.05 / 2.02 / 1.98
AliMeeting near/far
7.13/23.14
5.94/22.02
5.88/26.18
6.27/24.81
7.89/34.23
5.84 / 21.27
5.47 / 21.52
WenetSpeech net/meeting
4.96/6.16
3.74 / 3.62
4.11/4.54
4.13/4.23
6.63/8.52
3.91 /4.31
3.74 / 4.16
FLEURS-zh
2.95
2.00
2.40
2.13
1.85
1.84
1.84
Table 1: ASR Performance (CER/WER, %) on Public Benchmarks. Bold and underlined values mark the best and second-best results in each row, respectively.
TestSet
Qwen3 ASR Flash
FunASR Flash
Doubao ASR 2.0
Seed 2.0 Lite
StepAudio 3 ASR Max
Hy ASR 3.0 Preview
Hy ASR 3.0
Chinese
4.94
4.15
3.78
5.56
5.30
3.35
3.18
English
8.76
7.54
6.42
8.78
10.68
5.78
5.50
Cantonese
6.00
19.40
7.49
7.90
7.00
4.92
4.81
Code-Switching
8.80
5.30
6.49
6.37
6.90
4.35
4.20
Table 2: ASR Performance (CER/WER, %) on Internal Benchmarks. Bold and underlined values mark the best and second-best results in each row, respectively.
Figure 3: Examples of intra-utterance disambiguation. Panels (a–b) illustrate semantic disambiguation, while panels (c–d) illustrate explicit verbal clarification. All contextual cues are contained within the spoken utterance. Model outputs are shown as target-span excerpts.
Testset
Qwen3 ASR Flash
FunASR Flash
Doubao ASR 2.0
Seed 2.0 Lite
StepAudio 3 ASR Max
Hy ASR 3.0 Preview
Hy ASR 3.0
Intra-utterance context
46.60
51.16
51.86
54.64
50.75
58.12
61.88
Dialogue QA context
71.68
77.33
78.47
82.20
69.32
79.44
83.94
Table 3: Context Utilization Evaluation (Recall, %). Bold and underlined values mark the best and second-best results in each row, respectively.
Testset
Qwen3 ASR Flash
FunASR Flash
Doubao ASR 2.0
Seed 2.0 Lite
StepAudio 3 ASR Max
Hy ASR 3.0 Preview
Hy ASR 3.0
Long audio
3.86
3.98
5.60
4.99
3.58
2.96
2.35
Child & Elderly
9.98
7.87
6.43
10.49
9.90
7.18
6.25
Low-volume & Noisy
14.72
9.27
15.75
17.31
43.02
8.48
7.49
Table 4: ASR Robustness Evaluation (CER, %). Bold and underlined values mark the best and second-best results in each row, respectively.
Figure 4: Human evaluation of dialect recognition.