A Real-Time Voice Activity Detection Based On Lightweight Neural

语音活动检测（VAD）是指在音频流中检测语音的任务，由于真实环境中存在众多未知噪声和低信噪比，这是一项具有挑战性的任务。最近，基于神经网络的VAD在一定程度上缓解了性能下降的问题。然而，现有研究大多采用过于庞大的模型并加入了未来的上下文，而忽略了模型的操作效率和延迟。本文提出了一种轻量级和实时的神经网络MagicNet，它利用了非正式和深度可分离的1-D卷积和GRU。我们的提议模型不依赖于未来的特征作为输入，与两种最先进的算法在合成的领域内和领域外的测试数据集上进行了比较。评估结果表明，MagicNet可以在更少的参数成本下实现更好的性能和鲁棒性。

MagicNet: A Lightweight and Real-Time Neural Network for Voice Activity Detection

MagicNet is a lightweight and real-time neural network that utilizes casual and depth separable 1-D convolutions and GRU to detect speech in an audio stream without relying on future features as input.

The proposed model is compared with two state-of-the-art algorithms on synthesized in-domain and out-domain test datasets and achieves improved performance and robustness with fewer parameter costs. The experiments show that MagicNet is efficient and has low latency. The paper also provides open-source code for the model.

Recent studies in this field have employed excessively large models and incorporated future context, while MagicNet does not rely on future features and is lightweight. Some related papers are 'A Lightweight and Accurate Voice Activity Detection Model for Wearable Devices Using Convolutional Neural Networks' and 'Real-Time Voice Activity Detection Using Deep Recurrent Neural Networks with Bidirectional Long Short-Term Memory'.