neural-beamformer 0.1.0

Hybrid Neural Guided Beamforming for Edge Devices

Neural-Guided Beamformer

Real-Time Speech Enhancement via Neural Priors and a Beamformer backend

---

📖 Overview

This is the official implementation of a hybrid audio enhancement system, for real-time operation on dual-microphone edge devices (with smartphone like form-factors).

The pipeline comprises two main stages:

1. Neural Priori: A Speech-Adapted Efficient Group Enhanced UNet estimates a probabilistic Ideal Ratio Mask (IRM).
2. Spatial Filtering: A beamforming backend, which iteratively optimizes spatial filters using the neural priors.

---

✨ Key Features & Performance

• Robust to Reverberation: Outperforms state-of-the-art complex spectral mapping models. Benchmarks available.
• Strictly Distortionless: Unlike purely time-domain non-linear synthesizers (like ConvTasNet) that suffer from spectral degradation, the spatial filtering stage guarantees the target speech remains unaltered.
• Low-Latency and Edge compatible

---

📂 Repository Structure

Directory	Description
python/	Core Python Implementation: Clean inference scripts, metrics evaluation, and command-line interfaces. Start here for desktop usage.
AudioEnhancerCpp/	Android Application: Complete Android Studio project (app/). Features a native C++ audio engine built via CMake and JNI.
models/	Weights: Pre-trained .pth and .onnx models used by the inference pipelines.
research/	Development & Training: Contains PyTorch Lightning training code, MATLAB prototyping scripts, and evaluation logic.

---

🚀 Getting Started

🐍 Python (Desktop / Server)

1. Navigate to the Python directory and install dependencies:

   cd python
   pip install -r requirements.txt
   

2. Run the enhancement inference on a noisy audio file:

   python -m src.inference --input data/noisy_sample.wav --output data/clean_output.wav
   

📱 Android (Mobile App)

The mobile application is built to run the enhancement entirely on-device using C++ for low-latency NPU/CPU execution.

1. Open Android Studio and select Open an Existing Project.
2. Select the AudioEnhancerCpp/ folder.
3. Allow Gradle to sync and fetch the required NDK version.
4. Build and run on a physical Android device or emulator (API 24+).

---

📄 License & Citation

This project is licensed under the MIT License.

To use this code in your research, please cite our work as:

@misc{hybrid_neural_beamforming_2026,
  author = {{Audio Signal Processing Intelligence and REsearch Labs (ASPIRE), Indian Institute of Information Technology Design and Manufacturing Kurnool}},
  title = {Hybrid Neural Guided Beamforming for Edge Devices},
  year = {2026},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/Senpai-sama06/neural-guided-beamformer}}
}