Speaker ID in
real-world noise.
A 17M-parameter speaker identification model evaluated on clean speech, environmental noise, and music: 3.11% overall EER, versus 5.06% for TitaNet Large.
One enrolled voice
Environmental noise and music
- Task
- Speaker ID
- Model size
- 17 million parameters
- Design target
- Real-world noisy speech
01
Model design
Noisy speech
Optimized for recordings with environmental noise and music.
17M parameters
A compact model, available through the AirCaps AI API on request, at $0.02 per audio hour.
Three test variants
A clean baseline plus environmental noise and music.
02
Evaluation
Results across clean and noisy test sets.
We score speaker ID as speaker verification, using equal error rate (EER), the error rate at which false acceptances equal false rejections, on VoxCeleb1-O utterances. Noisy variants mix each utterance with MUSAN environmental noise or music at −10 and −5 dB SNR, so the noise is louder than the speech.
Speaker identification benchmark
Overall · EER (%) · Lower is betterEqual error rate (EER). Noise results include only cases where noise is louder than speech: −10 and −5 dB source SNR. Overall is the unweighted mean of clean plus four noisy conditions.
Full results table
| Model | Parameters | Overall | Clean | Environmental noise | Music |
|---|---|---|---|---|---|
| AirCaps Speaker ID | 17M | 3.11% | 0.41% | 3.53% | 4.03% |
| TitaNet Large | 25.3M | 5.06% | 0.82% | 4.12% | 8.11% |
| ECAPA-TDNN | 20.8M | 5.12% | 0.86% | 4.53% | 7.84% |
| ERes2NetV2 | 17.8M | 13.59% | 6.39% | 14.11% | 16.66% |
| WavLM Base+ SV | 101.1M | 18.57% | 4.72% | 21.52% | 22.55% |
03
Dataset
Hugging Face
Evaluation dataset
The VOX1O-MUSAN dataset accompanying this evaluation includes clean, environmental noise, and music variants, plus an overall aggregate. The Hugging Face release contains the audio, labels, and test conditions used for the reported results.
Model access