How microsoft/dns-challenge works
Project Type and Purpose
• This is a data-processing and evaluation pipeline for the ICASSP 2023 Deep Noise Suppression (DNS) Challenge.
• It automates:
• Downloading raw speech, noise, and RIR datasets
• Synthesizing noisy/clean speech pairs for training
• Evaluating enhancement models via DNSMOS metrics and Word Accuracy (WAcc)
A. Data Acquisition
1. Download scripts (bash files download-dns-challenge-5-*.sh)
2. Destination folder datasets_fullband/ (data store)
B. Data Synthesis
1. Configuration file (noisyspeech_synthesizer.cfg)
2. Synthesizer script (noisyspeech_synthesizer_singleprocess.py or pdns_noisyspeech_synthesizer_singleprocess.py)
3. Support modules (audiolib.py, utils.py)
4. Unit tests (unit_tests_synthesizer.py)
C. Evaluation
1. DNSMOS models (/DNSMOS folder with ONNX files)
2. pDNSMOS (personalized DNSMOS)
3. WAcc script (WAcc/WAcc.py)
4. External services: SpeechBrain speaker-embedding model, torchaudio
Architectural Patterns and Principles
• Linear pipeline: Acquisition → Synthesis → Validation → Evaluation
• Clear separation of concerns: each stage is a standalone script/module
• Config-driven: behavior of synthesizer is controlled by *.cfg file
Technologies and Frameworks
• Python 3.6+
• Audio I/O: soundfile, librosa, torchaudio
• Model inference: ONNX runtime (for DNSMOS)
• Speaker embeddings: SpeechBrain (speechbrain/spkrec-ecapa-voxceleb)
• Unit testing: standard Python unittest
Step-by-Step Flowchart Outline
[Download Scripts]
↓
[datasets_fullband/ (raw speech, noise, RIR)]
↓
[noisyspeech_synthesizer_singleprocess.py (uses cfg, audiolib, utils)]
↓
[Noisy/Clean Pairs Output Folder]
↙ ↘
[unit_tests_synthesizer.py] [DNSMOS ONNX Models]
↘ ↙
[Validation OK] [Quality Scores (SIG, BAK, OVRL)]
↓
[WAcc/WAcc.py] → Word Accuracy