urgent2025_challenge

Dataset Overview

• Dataset Name: urgent2025_challenge
• Purpose: Data preparation scripts for the URGENT 2025 Challenge.
• Compatibility: Generated metadata files are compatible with the baseline code.

Update Log

• 2024‑11‑27: Added troubleshooting guide for known issues.
• 2024‑11‑19: Modified ESTOI evaluation script for determinism.
• 2024‑11‑18: Added missing files required for Track 2 data preparation.
• 2024‑11‑16: Updated several data‑preparation and evaluation scripts.

Notes

• Validation Set: The generated validation set differs from the official one; the official set can be obtained here.
• Data Usage: The default data/speech_train subset is intended for dynamic mixing within the ESPnet framework.

System Requirements

• CPU Cores: >8
• Disk Space:
  • Track 1: ≥1.3 TB
  • Track 2: ??? TB
• Data Sizes:
  • Speech Data:
    • DNS5 speech: 318 GB
    • LibriTTS: 51 GB
    • VCTK: 12 GB
    • WSJ: 55 GB
    • EARS: 61 GB
    • CommonVoice 19.0 speech:
      • Track 1: 421 GB
      • Track 2: ??? GB
    • MLS:
      • Track 1: 120 GB
      • Track 2: ??? TB
  • Noise Data:
    • DNS5 noise: 93 GB
    • WHAM! noise: 76 GB
    • FSD50K: 30 GB
    • FMA: 60 GB
  • RIR Data:
    • DNS5 RIRs: 6 GB
  • Others:
    • Default simulated validation data: 2 GB
    • Simulated wind noise: 1 GB

Usage Instructions

1. Initialize Submodules:

   git submodule update --init --recursive
   

2. Install Environment:

   conda env create -f environment.yaml
   conda activate urgent2025
   

3. Obtain CommonVoice Dataset Link:
   • Retrieve the v19.0 download link from Common Voice.
4. Create Symbolic Links:
   • Create symlinks for wsj0 and wsj1 data.
5. Configure FFmpeg:
   • Modify the FFmpeg path in simulation/simulate_data_from_param.py.
6. Run Scripts:

   ./prepare_espnet_data.sh
   

7. Install eSpeak‑NG:
   • Required for phoneme similarity metric computation.

Troubleshooting

• MLS Extraction Error: Delete the failed file and rerun the script.
• FMA Processing Warning: Can be ignored.