ABSTRACT
On the one hand, the relationship between formant frequencies and vocal tract length (VTL) has been intensively studied over the years. On the other hand, the connection involving mel-frequency cepstral coefficients (MFCCs), which concisely codify the overall shape of a speaker's spectral envelope with just a few cepstral coefficients, and VTL has only been modestly analyzed, being worth of further investigation. Thus, based on different statistical models, this article explores the advantages and disadvantages of the latter approach, which is relatively novel, in contrast to the former which arises from more traditional studies. Additionally, VTL is assumed to be a static and inherent characteristic of speakers, that is, a single length parameter is frequently estimated per speaker. By contrast, in this paper we consider VTL estimation from a dynamic perspective using modern real-time Magnetic Resonance Imaging (rtMRI) to measure VTL in parallel with audio signals. To support the experiments, data obtained from USC-TIMIT magnetic resonance videos were used, allowing for the 2D real-time analysis of articulators in motion. As a result, we observed that the performance of MFCCs in case of speaker-dependent modeling is higher, however, in case of cross-speaker modeling, which uses different speakers' data for training and evaluating, its performance is not significantly different of that obtained with formants. In complement, we note that the estimation based on MFCCs is robust, with an acceptable computational time complexity, coherent with the traditional approach.
