Emotion and mental state recognition from speech

As research in speech processing has matured, attention has gradually shifted from linguistic-related applications such as speech recognition towards paralinguistic speech processing problems, in particular the recognition of speaker identity, language, emotion, gender, and age. Determination of a speaker’s emotion or mental state is a particularly challenging problem, in view of the significant variability in its expression posed by linguistic, contextual, and speaker-specific characteristics within speech. In response, a range of signal processing and pattern recognition methods have been developed in recent years. Recognition of emotion and mental state from speech is a fundamentally multidisciplinary field, comprising contributions from psychology, speech science, linguistics, (cooccurring) nonverbal communication, machine learning, artificial intelligence and signal processing, among others. Some of the key research problems addressed to date include isolating sources of emotion-specific information in the speech signal, extracting suitable features, forming reduced-dimension feature sets, developing machine learning methods applicable to the task, reducing feature variability due to speaker and linguistic content, comparing and evaluating diverse methods, robustness, and constructing suitable databases. Studies examining the relationships between the psychological basis of emotion, the effect of emotion on speech production, and the measurable differences in the speech signal due to emotion have helped to shed light on these problems; however, substantial research is still required. Taking a broader view of emotion as a mental state, signal processing researchers have also explored the possibilities of automatically detecting other types of mental state which share some characteristics with emotion, for example stress, depression, cognitive load, and ‘cognitive epistemic’ states such as interest, scepticism, etc. The recent interest in emotion recognition research has seen applications in call centre analytics, human-machine and humanrobot interfaces, multimedia retrieval, surveillance tasks, behavioural health informatics, and improved speech recognition. This special issue comprises nine articles covering a range of topics in signal processing methods for vocal source and acoustic feature extraction, robustness issues, novel applications of pattern recognition techniques, methods for detecting mental states and recognition of non-prototypical spontaneous and naturalistic emotion in speech. These articles were accepted following peer review, and each submission was handled by an editor who was independent from all authors listed in that manuscript. Herein, we briefly introduce the articles comprising this special issue. Trevino, Quatieri and Malyska bring a new level of sophistication to an old problem, detecting signs of depressive disorders in speech. Their measures of depression come from standard psychiatric instruments, Quick Inventory of Depressive Symptomatology and Hamilton Depression rating scales. These are linked to measures of speech timing that are much richer than the traditional global measures of speech rate. Results indicate that different speech sounds and sound types behave differently in depression, and may relate to different aspects of depression. Caponetti, Buscicchio and Castellano propose the use of a more detailed auditory model than that embodied in the widely employed mel frequency cepstral coefficients, for extracting detailed spectral features during emotion recognition. Working from the Lyon cochlear model, the authors demonstrate improvements on a five-class problem from the speech under simulated and actual stress database. Their study also further validates the applicability of long short-term memory recurrent neural networks for classification in emotion and mental state recognition problems. Callejas, Griol and López-Cózar propose a mental state prediction approach that considers both speaker * Correspondence: [email protected] School of Electrical Engineering and Telecommunications, The University of New South Wales, Sydney, NSW 2052, Australia Full list of author information is available at the end of the article Epps et al. EURASIP Journal on Advances in Signal Processing 2012, 2012:15 http://asp.eurasipjournals.com/content/2012/1/15