Optimal sensor placement in electromagnetic articulography recording for speech production study

The interrelationship between the face and vocal tract configuration during audiovisual speech

Speech perception is improved when we are able to see the person who is speaking, but how visual speech cues are used to improve speech perception is currently unclear. Brain imaging has revealed that regions responsible for motor control are active during the perception of speech, opening up the possibility that visual cues are mapped onto an internal representation of the vocal tract. Here, we show that there is sufficient information in the configuration of the face to recover the vocal tract configuration and that the key areas responsible for driving the correspondence vary in accordance with the articulation required to form the acoustic signal at the appropriate point in a sentence.

It is well established that speech perception is improved when we are able to see the speaker talking along with hearing their voice, especially when the speech is noisy. While we have a good understanding of where speech integration occurs in the brain, it is unclear how visual and auditory cues are combined to improve speech perception. One suggestion is that integration can occur as both visual and auditory cues arise from a common generator: the vocal tract. Here, we investigate whether facial and vocal tract movements are linked during speech production by comparing videos of the face and fast magnetic resonance (MR) image sequences of the vocal tract. The joint variation in the face and vocal tract was extracted using an application of principal components analysis (PCA), and we demonstrate that MR image sequences can be reconstructed with high fidelity using only the facial video and PCA. Reconstruction fidelity was significantly higher when images from the two sequences corresponded in time, and including implicit temporal information by combining contiguous frames also led to a significant increase in fidelity. A “Bubbles” technique was used to identify which areas of the face were important for recovering information about the vocal tract, and vice versa, on a frame-by-frame basis. Our data reveal that there is sufficient information in the face to recover vocal tract shape during speech. In addition, the facial and vocal tract regions that are important for reconstruction are those that are used to generate the acoustic speech signal.

Closed-set speaker conditioned acoustic-to-articulatory inversion using bi-directional long short term memory network

Estimating articulatory movements from speech acoustic representations is known as acoustic-to-articulatory inversion (AAI). In this work, a speaker conditioned AAI (SC AAI) is proposed using a bi-directional LSTM neural network, where training is performed by pooling acoustic-articulatory data from multiple speakers along with their corresponding speaker identity information. For this work, 7.24 h of multi-speaker acoustic-articulatory data are collected from 20 speakers speaking 460 English sentences. Experiments with 20 speakers indicate that the SC AAI model performs better than SD AAI model with an improvement of correlation coefficient by 0.036 (absolute) between the original and estimated articulatory movements.