ModelScope's FunASR Beats Whisper by 13x With a Single Model Call
FunASR's open-source toolkit runs 170x faster than Whisper, adds LLM-powered ASR, speaker diarization, and a drop-in OpenAI-compatible server.
Takeaways
- 170x realtime speed: SenseVoice-Small runs 13x faster than Whisper-large-v3 on GPU, and faster on CPU than Whisper manages on GPU.
- v1.3.3 ships a drop-in OpenAI API server:
funasr-server --device cudaexposes a fully OpenAI-compatible transcription endpoint at localhost:8000. - All-in-one pipeline: One model call handles VAD, ASR, punctuation, speaker diarization, and emotion detection — no extra libraries needed.
- LLM-powered Fun-ASR-Nano: Combines a SenseVoice encoder with Qwen3-0.6B decoder, trained on tens of millions of hours across 31 languages.
- MCP Server + vLLM: AI agents like Claude can now call FunASR directly; vLLM acceleration adds 2-3x throughput for the LLM-based model.
- Free and MIT licensed: Available via pip install funasr; models on Hugging Face and ModelScope.
FunASR, ModelScope's open-source speech recognition toolkit, has been surging on GitHub with over 700 new stars in a single week, crossing 17,400 total. The recent v1.3.3 release added a one-command server, an MCP integration for AI agents, and a vLLM-powered inference engine , turning what was already a fast transcription library into a full production-grade speech platform.
The problem with stitching together a pipeline
Anyone who has built a real transcription pipeline knows the tax: Whisper for ASR, pyannote for speaker diarization, a separate VAD model to chop long audio, another model for punctuation. With Whisper, you need to combine 5 separate projects for a production pipeline. FunASR collapses all of that into a single AutoModel call.
One model call handles VAD segmentation, speech recognition, punctuation, and speaker diarization automatically. The output is structured text with speaker labels and millisecond timestamps , no glue code required.
Speed that changes the hardware calculus
The headline number is 170x realtime on GPU for SenseVoice-Small. That sounds like marketing, but the benchmark is reproducible: 184 long-form audio files totaling 192 minutes, with the key takeaway being that FunASR models run on CPU faster than Whisper runs on GPU.
Here is how the models stack up:
| Model | GPU Speed | CPU Speed | vs Whisper-large-v3 |
|---|---|---|---|
| SenseVoice-Small | 170x realtime | 17x realtime | 13x faster |
| Paraformer-Large | 120x realtime | 15x realtime | 9x faster |
| Fun-ASR-Nano | 17x realtime | 3.6x realtime | 1.3x faster |
| Whisper-large-v3-turbo | 46x realtime | Not viable | 3.4x faster |
| Whisper-large-v3 | 13x realtime | Not viable | Baseline |
On FunASR's published benchmark, SenseVoice-Small runs at 170x realtime on GPU versus 13x realtime for Whisper-large-v3 , about 13x faster , and even runs at 17x realtime on CPU, faster than Whisper-large-v3 manages on a GPU. For teams processing large audio archives or running on cheap hardware, this is a genuine unlock.
Why it's so fast: parallel decoding
The speed advantage is architectural, not just engineering. Whisper is autoregressive , it generates one word at a time, each token depending on the previous one. Paraformer is a non-autoregressive (NAR) ASR model that generates the entire output in parallel, achieving significant speedups over autoregressive models like Whisper while maintaining competitive accuracy.