ASR Benchmark 2025: FunASR vs Whisper — Full Results

[LauraGPT]

Comprehensive ASR Benchmark

We have published a full benchmark comparing 7 ASR models on GPU and 3 on CPU, tested on 184 long-form Chinese audio files (192 minutes total).

Key Results (GPU, PyTorch)

Model	Type	Speed
SenseVoice-Small	NAR	170x realtime
Paraformer-Large	NAR	120x realtime
Whisper-large-v3-turbo	AR	46x realtime
Whisper-large-v3-turbo (FunASR)	AR	26x realtime
faster-whisper-large-v3	AR	21.5x realtime
Fun-ASR-Nano	LLM	17x realtime
Whisper-large-v3	AR	13.4x realtime

CPU Results

Model	Speed
SenseVoice-Small	17.2x realtime
Paraformer-Large	15.6x realtime
Fun-ASR-Nano	3.6x realtime
Whisper models	❌ Impractical on CPU

Links

• Full interactive benchmark page
• 中文版评测页面

Try it now

pip install --upgrade funasr

from funasr import AutoModel
model = AutoModel(model="FunAudioLLM/SenseVoiceSmall", hub="hf", vad_model="funasr/fsmn-vad", device="cuda")
result = model.generate(input="audio.wav")