Internal Projects

Learning muscle activation-acoustic map using a (deep) neural network

Recently biomechanical models of human speech production apparatus has been developed (www.artisynth.org). The purpose of this model is to study speech production and to understand relation between muscle activation patterns, articulation and acoustics. To achieve this purpose, lots of simulations needs to be done by using this model. Another alternative is to choose some limited number of patterns from muscle activation space, run the simulations and save the articulation and acoustic output. Then a neural network (NN) is utilized to learn the relation between two spaces and capability of NN is used to predict the articulatory or acoustic output for any muscle activation patterns. In this thesis, the biomechanical model will be used to generate training and test data sets which are used for training and evaluation the NN. Results will be analyzed in order to explore how speech production is planned and what are the limitations of this method. The results of this study could be published in a conference or journal.

Requirements of applicant: Knowledge in neural networks and speech technology, MATLAB, and Java programing
Suitable as: Master Project

Supervisor: Olov Engwall

Contact: Saeed Dabbaghchian

From vocal tract resonance frequencies to vocal tract area function

Human's speech production apparatus is a very complex system which has been studied by researchers of different fields and still lots of questions is unanswered. One aspect of speech is acoustics which study wave propagation in human's vocal tract. Vocal tract tube or area of cross-sections (area function) is analyzed to calculate resonance frequencies. In some applications, we need to solve the inverse problem by estimating the area function for desired resonance frequencies. Based on Fant's perturbation theory, a desired formants can be achieved using an iterative method. An alternative to this method one could generate samples of area functions and calculate the corresponding formants. A machine learning method (e.g. neural network) is utilized to learn the relationship between area function and formants. Generalization capability of the algorithm may be used to predict the area function for any unseen area function. The results of this study could be published in a conference or journal.

Requirements of applicant: Knowledge in speech technology and machine learning, MATLAB, and Java programing
Suitable as: Master Project

Supervisor: Olov Engwall

Contact: Saeed Dabbaghchian