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

[LauraGPT]

Important

Accuracy update (2026-07-15): This post previously mixed results from two benchmark paths and attributed pipeline capabilities to individual checkpoints. The tables below now preserve the published scopes separately and use the current model boundaries.

General ASR benchmark

The public comparison uses 184 long-form Chinese audio files, totaling 11,539 seconds (192.3 minutes), on one NVIDIA H100 80GB HBM3 GPU. Quality is reported as character error rate (CER); throughput is reported as times realtime (RTFx = audio duration / processing time). Lower CER and higher RTFx are better.

Model	Device	Speed	CER	Scope note
SenseVoice-Small	GPU	169.6x	7.81%	ASR plus language, emotion, and audio-event tags; tags stripped before CER.
Paraformer-Large	GPU	119.6x	10.18%	Chinese ASR used with the FunASR VAD/punctuation pipeline.
Fun-ASR-Nano	GPU	17.0x	8.06%	Standard LLM-ASR path in this table, not the separate vLLM batch result.
GLM-ASR-Nano	GPU	37.1x	31.07%	Multilingual LLM-ASR path.
Whisper-large-v3-turbo (OpenAI)	GPU	46.1x	21.71%	OpenAI Whisper implementation.
Whisper-large-v3 (OpenAI)	GPU	13.4x	20.02%	Baseline Whisper row.
SenseVoice-Small	CPU	17.2x	7.81%	CPU run from the same benchmark workflow.
Paraformer-Large	CPU	15.6x	10.18%	CPU batch path.
Fun-ASR-Nano	CPU	3.6x	8.06%	Heavier LLM-ASR path, still above realtime in this run.

These are benchmark results for the stated data, hardware, runtime, batching, and timing scope. They are not latency or throughput guarantees for other deployments.

Separate Fun-ASR-Nano vLLM benchmark

The vLLM guide reports a separate 184-file set totaling 11,541 seconds. Keep these rows separate from the general table above:

Fun-ASR-Nano runtime	RTFx	CER
PyTorch baseline	21x	8.06%
vLLM batch	340x	8.20%
Offline service, speaker diarization disabled	102x	8.14%
Offline service, speaker diarization enabled	46x	8.19%

The 340x result is offline vLLM batch throughput. It should not be read as streaming latency or substituted into the general benchmark table without preserving its runtime and batching scope.

Capability boundaries

• SenseVoiceSmall supports ASR for Mandarin, Cantonese, English, Japanese, and Korean, together with utterance-level language, emotion, and audio-event tags.
• The base Fun-ASR-Nano-2512 checkpoint covers Chinese, English, Japanese, seven Chinese dialect groups, and 26 regional accents. The separate Fun-ASR-MLT-Nano-2512 checkpoint is the 31-language model.
• Speaker diarization is supplied by a separate speaker model such as CAM++ in a composed FunASR pipeline; it is not built into SenseVoiceSmall or Fun-ASR-Nano.
• VAD, punctuation, timestamps, hotwords, and streaming availability depend on the selected model and runtime path. For example, the Paraformer WebSocket service is the mature streaming route; a checkpoint should not be labeled streaming merely because the toolkit has a streaming server.
• FunASR toolkit source is MIT-licensed. Model weights follow their respective model cards and may use different terms.

Minimal smoke test

pip install -U torch torchaudio funasr

from funasr import AutoModel

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

Reproduce and compare

• General benchmark report
• RTFx definitions and reproducibility fields
• Fun-ASR-Nano vLLM guide
• Migration benchmark script
• Model selection guide

For a meaningful comparison on your own data, record the model revision, runtime, hardware, batch size, VAD/punctuation/speaker settings, warmup policy, timing scope, failed files, and CER/WER normalization. A single clean demo clip is not a deployment benchmark.

[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.