Benchmark: FunASR 170x realtime vs Whisper 13x — non-autoregressive alternative
For those looking for faster alternatives to Whisper, FunASR offers non-autoregressive models that are significantly faster:
| Model | Speed (GPU) | Speed (CPU) | Languages |
|---|---|---|---|
| FunASR SenseVoice | 170x realtime | 17x realtime | 5 |
| FunASR Paraformer | 120x realtime | 15x realtime | 2 |
| Whisper-large-v3 | 13x realtime | ❌ | 57 |
Key advantage: FunASR runs on CPU faster than Whisper runs on GPU.
Also includes built-in speaker diarization and emotion detection.
pip install 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")
Benchmark details: https://modelscope.github.io/FunASR/benchmark.html
Useful table. Worth naming which speed it measures though: 170x realtime is batch throughput, which is exactly what you want for working through an archive. For a live conversation the number that matters is how long after someone stops speaking you get the final token, and non-autoregressive models do not automatically win there.
Worth separating throughput from latency when reading these numbers. A 170x realtime factor is a batch measurement, and in a streaming setting what a user feels is time to first token plus how often the model revises text it has already emitted. Both matter, and a single speed multiple hides them.