README.md

metadata

title: HybridGAN Speech Bandwidth Extension
emoji: 🎧
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.9.0
app_file: app.py
pinned: false
license: mit

🎧 HybridGAN-BWE: Speech Bandwidth Extension & Telephony Restoration

Band-limited narrowband telecommunication signals (standard landline G.711 codecs, mobile GSM channels, handset microphone cut-offs) restrict speech bandwidth to 300 Hz – 3400 Hz or 0 Hz – 4000 Hz. This severely degrades voice naturalness, intelligibility, and speaker identification.

HybridGAN-BWE presents a real-time, parameter-efficient Speech Bandwidth Extension system that reconstructs high-fidelity wideband speech (16 kHz) from narrowband input (8 kHz).

---

🚀 Key Features

• Dual-Path Spectral Encoder: Combines 2D spatial spectrogram convolutions and 1D temporal frame features fused via Multi-Head Cross-Attention.
• Domain Randomization: Robust against real-world telephone noise, GSM/G.711 compression, and handset high-pass cut-offs.
• Parameter Efficient: Squeezed 2D Spectrogram Discriminators (channels reduced from 32 to 16, saving 67% discriminator parameters).
• CPU Compatible: Optimized for CPU inference on Hugging Face Free Tier Spaces.

---

📦 Hugging Face Space Repository Structure

.
├── app.py              # Main Gradio application entrypoint
├── README.md           # Space Metadata & Documentation
├── requirements.txt    # Minimal dependencies (PyTorch, Gradio, Librosa)
├── config.yaml         # Model configuration parameters
├── best_model.pth      # Lightweight Generator model checkpoint
├── models/             # Model generator and discriminator definitions
├── modules/            # Spectral branch, waveform branch, and attention modules
└── utils/              # Config loader utility

---

🛠️ Local Testing Guide

To test this space locally on your machine:

cd hf_space
pip install -r requirements.txt
python app.py

Open http://127.0.0.1:7860 in your browser.