Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
benchmark: expand to 8 GPU models + 3 CPU models with full comparison
游雁18344cde5b16
diff --git a/README.md b/README.mdindex 552e6d376..7bbccf28b 100644--- a/README.md+++ b/README.md@@ -34,23 +34,39 @@<a name="benchmark"></a>-## Benchmark: FunASR vs Whisper (NVIDIA H100)+## Benchmark: Comprehensive ASR Comparison (NVIDIA H100)++> 8 models on GPU, 3 models on CPU. 184 long-form Chinese audio files (192 min total). [Full report](https://modelscope.github.io/FunASR/benchmark.html)++### GPU Results++| Model | Type | Speed | CER | Highlights |+|-------|------|-------|-----|-----------|+| **SenseVoice-Small** | NAR | **170x** realtime | **8.92%** | Fastest + most accurate |+| **Paraformer-Large** | NAR | **120x** realtime | 12.71% | Best speed/accuracy tradeoff |+| Whisper-large-v3-turbo (OpenAI) | AR | 46x realtime | 21.71% | Fastest Whisper variant |+| **GLM-ASR-Nano** | LLM | 37x realtime | 31.07% | Multi-language LLM-based |+| Whisper-large-v3-turbo (FunASR) | AR | 26x realtime | 22.70% | With VAD pipeline |+| faster-whisper-large-v3 | AR | 21.5x realtime | 17.00% | CTranslate2 optimized |+| **Fun-ASR-Nano** | LLM | 17x realtime | 10.56% | Best LLM-based accuracy |+| Whisper-large-v3 (OpenAI) | AR | 13.4x realtime | 20.02% | Baseline |++### CPU Results++| Model | Speed | CER | Note |+|-------|-------|-----|------|+| **SenseVoice-Small** | **17.2x** realtime | 5.14% | Production-ready on CPU |+| **Paraformer-Large** | **15.6x** realtime | 9.30% | Production-ready on CPU |+| **Fun-ASR-Nano** | 3.6x realtime | 7.60% | Still faster than realtime |+| Whisper models | - | - | Impractical on CPU (>2h for 192min audio) |-> Tested on 184 long-form Chinese audio files (192 min total). [Full report](https://modelscope.github.io/FunASR/benchmark.html)-| Model | Speed | CER | vs Whisper-large-v3 |-|-------|-------|-----|--------------------|-| **SenseVoice-Small** (FunASR) | **212x** realtime | **8.92%** | 🚀 **15x faster, 55% lower error** |-| **Paraformer-Large** (FunASR) | **128x** realtime | 12.71% | 🚀 **9x faster, 37% lower error** |-| faster-whisper-large-v3 | 21.5x realtime | 17.00% | 1.6x faster |-| Whisper-small (OpenAI) | 37.6x realtime | 26.79% | Faster but much worse accuracy |-| Whisper-large-v3 (OpenAI) | 13.7x realtime | 20.02% | Baseline |```python-# 3 lines to beat Whisper+# 3 lines to beat Whisper - 170x realtime, 55% lower error ratefrom funasr import AutoModelmodel = AutoModel(model="iic/SenseVoiceSmall", vad_model="fsmn-vad", device="cuda")-result = model.generate(input="meeting.wav") # 15x faster, better accuracy+result = model.generate(input="meeting.wav")```<a name="whats-new"></a>