Emotion detection from voice recordings

点击翻译按钮获取中文摘要

The speech signal carries several levels of information. From a voice recording, we can extract pronounced words, the speaker’s identity, the spoken language or even the speaker’s emotional state. In the last five years, there has been a great increase of interest in the field of emotion recognition based on different human modalities, such as speech, heart rate, etc. Building a robust emotion detection system can be very useful in several areas such as medicine and telecommunications. In the medical field, detecting emotions in general can be considered an important tool for diagnosing and following patients suffering from depression. The identification of the emotional state of a person from his voice opens new perspectives for the development of an automated dialogue system, one capable of communicating with patients at home daily and even several times a day to produce a report for the physician. Although much research has been done, actual emotion recognition systems performances are still not adequate for these real life applications. The majority of emotion detection systems have been designed by focussing on the speech modeling rather than the feature extraction aspect. Usually, feature vectors are made of a combination of classical cepstral features (such as MFCC) augmented with some prosodic characteristics (such as speech intonation). The modest results presented in the literature make us think that we should focus on improving the quality of features used for training emotion identification systems instead of focussing on the modeling aspect. For this reason, our research will focus on extracting more accurate features from the speech signal. Three main axes will be established in order to reach this objective. The first axe consists of working on the development of a DNN architecture capable of learning the classification function directly from the raw speech signal. The second axe consists of developing another DNN architecture capable of learning a normalization function in order to be able to mix different databases made of speech signals from different recording conditions. Finally, in the third axe, we will concentrate on high-level features such as phoneme durations and other evidence extracted from words said by the speaker. Improvements made during this research will be very practical for building reliable healthcare applications. Indeed, such applications would be useful for health professionals by helping them improve the quality of treatments in Canada. Moreover, our findings could be applied to several other fields of speech technologies. The approach presented would facilitate the development of command and control applications without the use of a complex and time/power consuming speech recognition engine. This would help start-ups or small businesses in Canada to add voice-based control to their applications without needing to hire a speech recognition specialist.