Accent is generally defined as a set of some typical pronunciation traits. These traits, when perceived, contribute to classify the speech either as non standard, or as coming from a specific (regional, sociological) variety. It has been shown that accent classification is a very difficult task, both for humans as well as for automatic classifiers. Many facets of what makes an accent remain still to be uncovered and the proposed subject aims at contributing to this aim. What about simulated accents? What if speakers exagerate pronunciation traits to mimic some other speaker or some non-standard variety? Speakers may be more or less gifted to play with accents, mimic various accents and even different voices. This raises important issues with respect to fundamental questions concerning the nature of accents and the characterisation of human voices. On an application side, automatic detection of accent simulation will contribute to security applications such as impostor detection.
Impostor detection is a very active area in the field of automatic speaker recognition. Applied methods include GMMs (gaussian mixture models), HMMs (Hidden Markov Models), SVMs (Support Vector Machines). Acoustic features typically correspond to MFCCs (mel frequency cepstrum coefficients) and MFCC derivatives. Automatic accent identification is a relatively recent issue within the field of automatic language recognition. However the question of accent impostors has, to the best of our knowledge, not been addressed yet.
Improve our knowledge/understanding of accent (perception/pruduction) using automatic speech processing methods.
Data collection of accented speakers and accent impostors. Definition of acoustic/prosodic/pronounciation features. Speech synthesis of controlled accent stimuli. Automatic classification and perceptual experiments involving natural accent impostor stimuli as well as synthetic accent-graded stimuli.