Search

International audience; Ultrasound imaging of the tongue and videos of lips movements can be used to investigate specific articulation in speech or singing voice. In this study, tongue and lips image sequences recorded during singing performance are used to predict vocal tract properties via Line Spectral Frequencies (LSF). We focused our work on traditional Corsican singing " Cantu in paghjella ". A multimodal Deep Autoencoder (DAE) extracts salient descriptors directly from tongue and lips images. Afterwards, LSF values are predicted from the most relevant of these features using a multilayer perceptron. A vocal tract model is derived from the predicted LSF, while a glottal flow model is computed from a synchronized electroglottographic recording. Articulatory-based singing voice synthesis is developed using both models. The quality of the prediction and singing voice synthesis using this method outperforms the state of the art method.

Résumé long de la communication visible au lien suivant: http://www.qmu.ac.uk/casl/conf/ultrafest%5F2013/docs/AJaumard-Hakoun_1_Ultrafest.pdf; International audience; The i-Treasures project, which officially began on 1 February 2013, is a 12-partner FP7 project that proposes to use multi- sensor technology to capture, preserve, and transmit four types of intangible cultural heritage, referred to as 'use cases': rare traditional songs, rare dance interactions, traditional craftsmanship and contemporary music composition. Methodologies used will include body and gesture recognition, vocal tract modeling, speech processing and electroencephalography (EEG). The "Rare traditional songs" use case, which will be the focus of our work, targets Corsican "cantu in Paghjella", Sardinian "Canto a tenore", Mt. Athos Greek byzantine hymns and the recent "Human beat box" styles. The final objective of the "Rare traditional songs" use case is to capture vocal tract movements with sufficient accuracy to drive a real-time 2D or 3D avatar of the vocal tract, which will in turn play a crucial role in the transmitting of captured invisible cultural heritage to future generations. The acquisition sensors are: multi-sensor helmet, nose-mounted accelerometer, camera, ultrasound probe, Electroglottograph, Microphone and breathing belt. This system is described in this study, and also the vocal tract modeling, the preliminary data processing, and the extracting tongue contour method.

HE NON-INVASIVE magnetic resonance (MR) scanning of the human vocal tract volume enables to develop the three-dimensional (3D) computer models of the vocal tract. These 3D models are necessary for understanding the basic physical principles for the creation of human speech and voice as close to reality as possible. The primary volume models of the human acoustic supraglottal spaces created from the MR images can then be transformed into the 3D finite element (FE) models [1]. Such models are helpful for modeling the real clinical situation, such as influence of various inborn defects in human supraglottal spaces on speech and voice or simulations of various postsurgical states in patients [2]. The quality of the developed FE models has to be checked by a sufficiently accurate numerical simulation of the subject phonation during the NMR scanning and therefore the simultaneous acoustic recording of subject voice during the scan procedure is very important [3], [4]. Head probe coils are commonly produced for the magnetic resonance imaginer (MRI) equipment working with strong magnetic field, but these systems produce a lot of unwanted acoustic noise. The solutions to the acoustic noise problems are low field MR scanners, but these are not usually provided with the head/neck coils. Therefore, this study focused on the development of a new MR receiving head coil, for a tomograph with low magnetic field to be used primarily in human vocal tract imaging.