Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 2 min read

Announcing wavewash (v0.2.0) on PyPI β€” Production Audio QC Engine with Local Web UI & PyTorch API

πŸš€ Excited to announce the official release of wavewash (v0.2.0) on PyPI! If you train Speech or Audio AI models (Whisper, StyleTTS, AudioCraft, MusicLM), your models inherit the acoustic flaws of their training data: πŸ‘‰

πŸš€ Excited to announce the official release of wavewash (v0.2.0) on PyPI!

If you train Speech or Audio AI models (Whisper, StyleTTS, AudioCraft, MusicLM), your models inherit the acoustic flaws of their training data:

πŸ‘‰ Clipped waveforms, DC offset, dead channels, and inconsistent loudness.

Saturated audio creates harsh distortion in neural vocoders. DC displacement burns dynamic headroom. Disjointed LUFS levels destabilize loss convergence, while silent channels corrupt stereo spatialization. Hand-cleaning thousands of hours of audio is impossible, and generic black-box cleaners introduce hallucinated artifacts.

WaveWash solves this: an open-source Python library, CLI, and embedded Web UI engineered for deterministic audio quality control, verification, and DSP repair.

πŸ” What Does WaveWash Do?

  1. 🩹 Verified DSP De-Clipping: Mathematically reconstructs clipped peaks via cubic splines, cutting distortion by 71.8% while guarding against irrecoverable saturation.
  2. 🎚️ DC Offset & Channel Repair: Eliminates DC bias, fixes dead stereo channels, and normalizes integrated loudness to broadcast-standard ITU-R BS.1770 LUFS.
  3. 🎯 Purpose-Driven Profiles: Pre-tuned with strict acoustic thresholds for asr, tts, speaker_id, and music_gen (or customize your own).
  4. 🐍 Python In-Memory API: Embeds directly into PyTorch Dataset loaders and streaming pipelines (analyze_file, apply_fix_pipeline) for zero-disk-overhead training.
  5. πŸ–₯️ Embedded Web Dashboard: Run wavewash app to launch a pre-compiled local web UI with interactive audio auditing and before/after players.
  6. πŸ›‘οΈ Non-Destructive: Never overwrites original audio. Clean files pass to clean/; rejected files route to categorized reason folders (low_snr/, clipping_heavy/).

πŸ“Š Real-Data Benchmarks

Validated across extensive speech and studio multi-track benchmarks:

βœ… Clipping Repair Rigor: 71.8% MSE distortion reduction on real clipped speech (LibriSpeech).
βœ… Zero False Rejections: 100% clean pass on true studio stems from MUSDB18-HQ.
βœ… Noise Floor Defense: High-accuracy SNR estimation and dual-tier clipping guardrails.
βœ… Multi-Core Speed: Processes multi-gigabyte audio corpora with multi-threaded parallel workers.

πŸ“¦ Quick Start & Links

Install directly from PyPI:
pip install wavewash

Launch the local interactive Web UI:
wavewash app --port 8420

Batch clean a dataset via CLI:
wavewash run --input ./raw_speech --output ./clean_speech --purpose asr

πŸ”— PyPI: https://pypi.org/project/wavewash/
πŸ’» GitHub: https://github.com/BoobalanArjunan/wavewash
πŸš€ 1-Click Colab: https://colab.research.google.com/github/BoobalanArjunan/wavewash/blob/main/notebooks/quickstart.ipynb

If you train audio models or prepare speech datasets, give it a spin! ⭐

AudioML #MachineLearning #DeepLearning #SpeechAI #AudioAI #TTS #ASR #Python #OpenSource #DataQuality #AudioEngineering #AI

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.