Edge AI Speech Enhancement

Speech enhancement pipeline removing background noise from audio signals, optimized for edge devices.

What It Does

• Removes background noise from audio recordings
• Optimized for CPU-only inference (<100ms latency)
• Works on resource-constrained devices (earbuds, wearables, IoT)
• Converts to ONNX for edge deployment

Why It Matters

Speech enhancement is crucial for:

• Clear voice calls in noisy environments
• Voice assistants reliability
• Hearing aids and audio devices
• IoT and embedded systems

Key Features

• Spectral subtraction algorithm
• Spectrogram visualization
• ONNX conversion (75% size reduction)
• CPU-only real-time inference
• PyTorch + Librosa pipeline

Installation

pip install -r requirements.txt

Or:

pip install torch torchaudio librosa soundfile numpy matplotlib

Usage

Jupyter Notebook (Recommended)

jupyter notebook notebook_inference.ipynb

Python Script

python scripts/inference.py --input input_audio/noisy_speech.wav --output output_audio/clean_speech.wav

ONNX Conversion

python scripts/convert_to_onnx.py

Performance

Metric	Value
Inference Latency	<100ms
Model Size Reduction	75%
Platform	CPU-only
Target Devices	Earbuds, IoT, Wearables

Pipeline

Noisy Audio → STFT → Spectrogram → Spectral Subtraction → ISTFT → Clean Audio

Project Structure

edge-ai-speech-enhancement/
├── notebook_inference.ipynb     # Demo notebook
├── README.md                   # This file
├── requirements.txt             # Dependencies
├── LICENSE                     # MIT License
└── scripts/
    ├── inference.py            # Main enhancement
    ├── convert_to_onnx.py     # ONNX conversion
    └── quantize_model.py      # Model quantization

Skills Demonstrated

• Audio signal processing (FFT/STFT)
• Spectral subtraction algorithms
• Model optimization for edge devices
• ONNX conversion and quantization
• Low-latency inference

Related Projects

• Real-Time VAD
• Audio Source Separation

Together these show a complete audio AI pipeline!

---

GitHub: github.com/satzgits/edge-ai-speech-enhancement