Benchmark: FunASR vs Whisper — Real-World Performance on Chinese Meeting Audio

[LauraGPT]

TL;DR

Model	Speed (GPU)	CER (Chinese)	Speaker ID	Streaming	CPU-viable
FunASR SenseVoice	170x realtime	4.2%	✅ Built-in	❌	✅ 17x
FunASR Fun-ASR-Nano	340x (vLLM)	8.2%	✅ Built-in	✅ WebSocket	✅ 3.6x
Whisper-large-v3	13x realtime	9.8%	❌ Needs pyannote	❌	❌
Whisper-large-v3-turbo	46x realtime	10.1%	❌	❌	❌

Test Setup

• Audio: 184 files, 192 minutes total, Chinese meeting recordings
• Hardware: NVIDIA A100 80GB
• Metrics: Real-Time Factor (RTFx), Character Error Rate (CER)

Key Findings

1. Speed: FunASR is 13x faster than Whisper on GPU

SenseVoice-Small processes 10 seconds of audio in just 70ms. On CPU, FunASR models run at 17x realtime — faster than Whisper runs on GPU.

2. Chinese Accuracy: FunASR wins on dialects and accents

FunASR supports 7 Chinese dialects (Wu, Cantonese, Min, Hakka, Gan, Xiang, Jin) and 26 regional accents. Whisper struggles with non-Mandarin Chinese.

3. All-in-one Pipeline: No assembly required

With Whisper, you need to combine 5 separate projects for a production pipeline:

• Whisper (ASR)
• pyannote (speaker diarization)
• silero-vad (voice activity detection)
• deepmultilingualpunctuation (punctuation)
• Custom code (emotion detection — not available)

FunASR does all this in one API call:

from funasr import AutoModel
model = AutoModel(model=\"paraformer-zh\", vad_model=\"fsmn-vad\", punc_model=\"ct-punc\", spk_model=\"cam++\")
result = model.generate(input=\"meeting.wav\")
# Returns: text + timestamps + speaker IDs + punctuation

4. OpenAI-Compatible API: Zero-code migration

If you are already using OpenAI Whisper API or any compatible client, switching to FunASR requires zero code changes:

pip install funasr vllm fastapi uvicorn python-multipart
funasr-server --device cuda
# Now POST to http://localhost:8000/v1/audio/transcriptions

5. Self-hosted & Free

• MIT license, no API keys needed
• All data stays on your machine
• No per-minute charges

How to Try

pip install torch torchaudio
pip install funasr

from funasr import AutoModel
model = AutoModel(model=\"iic/SenseVoiceSmall\", device=\"cuda\")
result = model.generate(input=\"your_audio.wav\")
print(result[0][\"text\"])

Links

• Full benchmark report
• Migration guide from Whisper
• Website
• Colab Quickstart

[prashantweb3task-collab]

Thanks for sharing these benchmarks. The speed difference is especially impressive, and it's interesting to see FunASR outperform Whisper on Chinese meeting audio, particularly with dialect and accent support. The built-in speaker diarization, VAD, and punctuation also make the deployment story much simpler compared to assembling multiple separate components around Whisper.

For teams working with large volumes of meeting recordings, reducing infrastructure complexity can be just as valuable as improving transcription accuracy. It's also great to see more open-source options emerging in the speech recognition space.

I'm curious whether anyone has tested FunASR in production workflows for note-taking and meeting summaries. Tools such as VoiceToNotes.ai rely heavily on accurate speech-to-text pipelines, so improvements in speed and multilingual accuracy could have a big impact on real-world productivity applications.