Automatic Speech Recognition on AISHELL-2 (test_ios)
2.08CERFireRedASR2S LLM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FireRedASR2S LLMModel Size=8B+, Inference Mode=Offline2026.04 | 2.08 | — | |
| FireRedASR-LLMModel Size=8B+ (↑)2026.04 | 2.12 | — | |
| FireRedASR-LLM#Params=8.3B2025.01 | 2.15 | — | |
| Seed-ASR#Params=12B+2025.01 | 2.27 | — | |
| Step-Audio2-miniModel Size=8B+ (↑)2026.04 | 2.29 | — | |
| Step-Audio2 MiniModel Size=8B+, Inference Mode=Offline2026.04 | 2.29 | — | |
| Qwen3-Omni-InstModel Size=30B-A3B (↑)2026.04 | 2.31 | — | |
| Qwen3-Omni InstructModel Size=30B-A3B, Inference Mode=Offline2026.04 | 2.31 | — | |
| NIM4-ASRModel Size=2.3B, Inference Mode=Offline2026.04 | 2.43 | — | |
| OursModel Size=2.3B2026.04 | 2.45 | — | |
| NIM4-ASRModel Size=2.3B, Inference Mode=Stream2026.04 | 2.49 | — | |
| FireRedASR-AED#Params=1.1B2025.01 | 2.52 | — | |
| Qwen3-ASR-1.7BModel Size=2.0B (↓)2026.04 | 2.6 | — | |
| Qwen3-ASR 1.7BModel Size=2.0B, Inference Mode=Offline2026.04 | 2.6 | — | |
| Fun-ASR-nanoModel Size=0.8B (↓)2026.04 | 2.73 | — | |
| Fun-ASR NanoModel Size=0.8B, Inference Mode=Offline2026.04 | 2.73 | — | |
| Paraformer-large#Parameters=220M, AR/NAR=NAR, LM=w/o, batchsize=12023.05 | 2.85 | 0.0251 | |
| Paraformer-Large#Params=0.2B2025.01 | 2.85 | — | |
| SenseVoice-L#Params=1.6B2025.01 | 3.04 | — | |
| Qwen-Audio#Params=8.4B2025.01 | 3.1 | — | |
| GLM-ASR-nanoModel Size=1.5B (↓)2026.04 | 3.45 | — | |
| GLM-ASR NanoModel Size=1.5B, Inference Mode=Offline2026.04 | 3.45 | — | |
| Whisper-Large-v3#Params=1.6B2025.01 | 4.96 | — | |
| UMAModel Type=NAR, Self-conditioning=false, RTF=0.085, Params(M)=105.42023.09 | 5.3 | — | |
| UMA + self-conditionModel Type=NAR, Self-conditioning=true, RTF=0.098, Params(M)=110.42023.09 | 5.3 | — | |
| WeNet Conformer-U2++#Parameters=47.30M, AR/NAR=AR, LM=w/, batchsize=12023.05 | 5.35 | — | |
| Conformer#Parameters=46.25M, AR/NAR=AR, LM=w/, batchsize=12023.05 | 5.35 | 1.43 | |
| K2 Transducer#Parameters=80M, AR/NAR=AR, LM=w/o, batchsize=12023.05 | 5.56 | — | |
| ESPnet Conformer#Parameters=46.25M, AR/NAR=AR, LM=w/, batchsize=12023.05 | 5.7 | — | |
| Hybrid (Conformer) + beam searchModel Type=AR, Beam search=true, RTF=0.954, Params(M)=116.42023.09 | 5.7 | — | |
| Paraformer#Parameters=46.3M, AR/NAR=NAR, LM=w/o, batchsize=12023.05 | 5.73 | 0.0168 | |
| CIF + SANModel Type=NAR2023.09 | 5.8 | — | |
| Conformer#Parameters=46.25M, AR/NAR=AR, LM=w/o, batchsize=12023.05 | 5.83 | 0.21 | |
| Hybrid (Conformer)Model Type=AR, Beam search=false, RTF=0.205, Params(M)=116.42023.09 | 6.3 | — | |
| LASO-largeModel Type=NAR, Params(M)=80.02023.09 | 6.7 | — |