Android DeepFilterNet

This repository provides an Android implementation of the DeepFilterNet noise suppression model with JNI bindings.

Overview

DeepFilterNet is a state-of-the-art neural network architecture designed for real-time noise suppression in audio streams. This Android implementation allows developers to integrate high-quality noise removal capabilities into Android applications.

Features

• Real-time noise suppression for Android applications
• JNI bindings for efficient integration with native Android code
• Optimized for mobile performance

Library Requirements

• Android API level 21+ (Android 5.0 or higher)
• Android NDK r21+
• Gradle 7.0+

Audio Requirements

• Sample Rate: 48 kHz
• Audio Depth: 16-bit PCM
• Channels: Mono

Installation

Add the following to your app's build.gradle file:

dependencies {
    // Option 1: Use DeepFilterNet with the model bundled directly in the library.
    // Choose this if you prefer a self-contained solution and are okay with a larger app size.
    implementation 'io.github.kaleyravideo:android-deepfilternet:x.y.z'

    // Option 2: Use DeepFilterNet with the model downloaded on demand during initialization.
    // Choose this if you want to minimize initial app size and download the model only when needed.
    implementation 'io.github.kaleyravideo:android-deepfilternet-lazymodel:x.y.z'
}

The model size is ~8MB.

Usage

Basic Usage

// Initialize the DeepFilterNet instance.
val deepFilterNet = com.rikorose.deepfilternet.NativeDeepFilterNet(context)
// Set the noise attenuation level (dB).
deepFilterNet.setAttenuationLimit(30f)

..

// Get the DeepFilterNet's frame length.
val frameLength = deepFilterNet.frameLength.toInt()
// Allocate a new direct ByteBuffer with the given frame length to interact with the native code.
val byteBuffer = ByteBuffer.allocateDirect(frameLength).apply {
    order(ByteOrder.LITTLE_ENDIAN) // Set byte order to match DeepFilterNet's expectation
}

..

// Process the audio frame using the DeepFilterNet model.
// The 'processFrame' method modifies the buffer in-place.
deepFilterNet.processFrame(byteBuffer)
    
..

// Release resources when done.
deepFilterNet.release()

For a full example, click here.

Documentation

For detailed documentation on the API, see the documentation.

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

Credits

This project is built upon a fork of the original DeepFilterNet work by Hendrik Schröter.<br/> Further information regarding the optimization process of the DeepFilterNet model for running on a mobile device is available in this document.

Demo app audio samples

• Voice generated with Narakeet.
• Background noise from Soundjay.

Contact

For questions or support, please open an issue on the GitHub repository or contact the Kaleyra Video team at cis-eu.video.engineering@tatacommunications.com.