Phonos: Ultra-Compact Universal 44.1kHz Speech Restorer
Phonos is an ultra-compact (78.7 MB) quantized neural speech restoration system engineered for real-world forensic recovery, 44.1kHz bandwidth extension, acoustic dereverberation, and severe distortion repair.
Core Capabilities
- 44.1 kHz Studio Resynthesis: Synthesizes missing high-frequency harmonics up to 22.05 kHz from band-limited (4kHzโ8kHz) telephone and radio recordings.
- Acoustic Dereverberation: Deconvolves severe room impulse responses and echo ($T_{60} > 0.8\text{s}$).
- Waveform Declipping: Repairs flat-topped waveform saturation.
- Ultra-Compact Mixed Precision: 78.7 MB checkpoint (Mixed NF4/FP8) running in <950 MB VRAM at 11.8x real-time on consumer GPUs.
Quick Start
git clone https://github.com/jaadu611/phonos
cd phonos
pip install -r requirements.txt
python cli.py recording.wav -o restored_studio.wav
Attribution & Architecture
Phonos is built on the VoiceFixer framework by Haohe Liu et al. (Interspeech 2022). It contributes selective Mixed NF4/FP8 quantization, high-frequency harmonic presence excitation, and standalone zero-dependency inference.