Articulatory Comparison of L1 and L2 Speech for Mispronunciation Diagnosis

Accurate Synthesis of Dysarthric Speech for Asr Data Augmentation

“…The number of augmented sentences was ten times that of the original TORGO dataset. The synthesized speech was used for training the DNN-HMM ASR model, trained on fMLLR transformed features Baseline configuration files provided in the Pytorch-kaldi repository for common speech databases like TIMIT, Librispeech were used as reference and the final architecture was based on experimental results using a small number of training set speakers (Khanal, Johnson et al 2021). Our ASR models includes light bidirectional GRU (Ravanelli, Parcollet et al 2019) architecture, with five layers containing 1024 cells each, activated by Relu activation function and dropout of 0.2.…”

“…The architecture applies monophone regularization (Ravanelli, Brakel et al 2018). A multi-task learning procedure was applied using two SoftMax classifiers, one estimating context-dependent states and the second one predicting monophone targets (Khanal, Johnson et al 2021). For testing, a leave-one-speaker-out cross-validation procedure was applied across the original TORGO dataset.…”

Tongue Contour Tracking and Segmentation in Lingual Ultrasound for Speech Recognition: A Review

“…EMA [ 39 , 40 , 41 ] is useful to localize the movement within the vocal tract by using electromagnetic transmitter coils to track the position of the attached electromagnetic sensors on the tongue, lips, and jaw. EMA may provide either a 2D or 3D landmark localization in milliseconds, but the system operation is complex and uncomfortable to be used in all cases on a daily basis; it might be more usable for conducting clinical studies at research centres.…”