DPDFNet: Evaluation Set

A multilingual, long-form, low-SNR evaluation set introduced in the paper “DPDFNet: Boosting DeepFilterNet2 via Dual-Path RNN”.

At a glance

• 12 languages: arabic, dutch, french, german, hungarian, korean, polish, portuguese, russian, spanish, turkish, vietnamese
• 9 noise scenes: airport, car, office, pub, rain, restaurant, street, subway, train
• SNR (dB): 0 / 5 / 10
• Duration: ~2.5 minutes per clip (includes speech-free intervals)
• Size: 324 clips (~13.5 hours)

What you get

This repo is meant for evaluation, and it includes:

• Noisy mixtures
• Clean references
• Enhanced outputs from the models reported in the paper (ours and open-source)

Recommended use

Use this set to evaluate:

• Causal / streaming enhancement models on long recordings
• Robustness to silence → speech transitions, non-stationary noises, and multilingual speech
• Stability on clips that may contain multiple speakers within a single recording

Citation

If you use this evaluation set, please cite the DPDFNet paper:

@article{rika2025dpdfnet,
  title   = {DPDFNet: Boosting DeepFilterNet2 via Dual-Path RNN},
  author  = {Rika, Daniel and Sapir, Nino and Gus, Ido},
  year    = {2025}
}