Online Own Voice Detection for a Multi-Channel Multi-Sensor In-Ear Device

Voice activity detection (VAD) aims for detecting the presence of speech in a given input signal, and is often first step voice -based applications such as communication systems. In context personal devices, own (OVD) sub-task VAD, since it targets person wearing device, while ignoring other speakers interference signals. This article summarizes recent single multi-microphone, multi-sensor, hearing aids related VAD techniques. Then, wearable in-ear device equipped with multiple microphones an accelerometer investigated OVD task using neural network embedding long short-term memory (LSTM) layers. The picks up user’s signal through air well vibrations body. However, besides external sounds sensitive to non-speech vocal noises (e.g. coughing, yawning, etc.) movement noise caused by physical activities. A mixing model proposed produce databases noisy observations used training testing frame-by-frame method. best model’s performance further studied different recorded interference. An ablation study reports on sub-sets sensors. results show that approach robust towards both user motion generated loud suitable real-time operation achieves 90-96 % accuracy challenging use scenarios short 10 ms processing frame length.