No abstract

Distant emotion recognition (DER) extends the application of speech emotion recognition to the very challenging situation that is determined by variable speaker to microphone distances. The performance of conventional emotion recognition systems degrades dramatically as soon as the microphone is moved away from the mouth of the speaker. This is due to a broad variety of effects such as background noise, feature distortion with distance, overlapping speech from other speakers, and reverberation. This paper presents a novel solution for DER, addressing the key challenges by identification and deletion of features from consideration which are significantly distorted by distance, creating a novel, called Emo2vec, feature modeling and overlapping speech filtering technique, and the use of an LSTM classifier to capture the temporal dynamics of speech states found in emotions. A comprehensive evaluation is conducted on two acted datasets (with artificially generated distance effect) as well as on a new emotional dataset of spontaneous family discussions with audio recorded from multiple microphones placed in different distances. Our solution achieves an average 91.6%, 90.1% and 89.5% accuracy for emotion happy, angry and sad, respectively, across various distances which is more than a 16% increase on average in accuracy compared to the best baseline method.

No abstract

No abstract