Learning to Detect Dysarthria from Raw Speech

Abstract

Speech classifiers of paralinguistic traits traditionally learn from diverse hand-crafted low-level features, by selecting the relevant information for the task at hand. We explore an alternative to this selection, by learning jointly the classifier, and the feature extraction. Recent work on speech recognition has shown improved performance over speech features by learning from the waveform. We extend this approach to paralinguistic classification and propose a neural network that can learn a filterbank, a normalization factor and a compression power from the raw speech, jointly with the rest of the architecture. We apply this model to dysarthria detection from sentence-level audio recordings. Starting from a strong attention-based baseline on which mel-filterbanks outperform standard low-level descriptors, we show that learning the filters or the normalization and compression improves over fixed features by 10% absolute accuracy. We also observe a gain over OpenSmile features by learning jointly the feature extraction, the normalization, and the compression factor with the architecture. This constitutes a first attempt at learning jointly all these operations from raw audio for a speech classification task.

Glottal Source Information for Pathological Voice Detection

“…For UA-Speech and TORGO, the database is split in order to maintain a good partition of speakers with different severities or intelligibility scores between the training, validation, and test sets, without having any overlap in speakers between the different sets. A similar method of data splitting has been used, for example, in [38]. Tables 3 and 4 show the data partition in UA-Speech and TORGO, respectively.…”

“…In order to develop deep learning models, existing studies have mainly used combinations of convolutional neural network (CNN) and multilayer perceptron (MLP) [31], [33]- [37]. In addition, some studies have explored combining CNN and long short-term memory (LSTM) networks [32], and combining LSTM and MLP [38] for detection of pathological voice from healthy speech. Even though different deep learning architectures have been studied in the recent pathological voice detection studies listed above, a systematic comparison between latest end-to-end methods and systems based on the traditional pipeline is still lacking in the study area.…”

“…In the end-to-end system framework, the glottal source has been taken advantage of in detection of emotion [50] and depression [47], and in TTS [39]. In the area of pathological voice detection, there is typically a relatively small amount of training data available (see, e.g., [38], [51]) in comparison to speech recognition [52] and speaker verification [53] systems that might use even thousands of hours of speech data. In order to avoid the data scarcity problem, some previous studies have used data augmentation techniques to artificially increase the amount of pathological voice training data [54], [55].…”

Automatic Assessment of Dysarthric Severity Level Using Audio-Video Cross-Modal Approach in Deep Learning

“…Moreover, certain degenerative illnesses, such as amyotrophic lateral sclerosis (ALS), often cause damages to various areas of the nervous system [27] [28]. Some researchers also indicate that the disease can also have influences on other areas, such as early childhood recognition, sensory defect, and intellectual disability [29] [30][31] [32]. However, depends on the severity levels of the damage to different areas, we can still define the leading root cause of the disease, and based on that we can group dysarthria into different types.…”

Prosody-Based Measures for Automatic Severity Assessment of Dysarthric Speech

“…Detecting dysarthria involves extracting hand-crafted acoustic features and using those features as inputs to a machine learning-based classifier [18][19][20]. Deep learning approaches are also possible where the raw speech signal or a set of elementary features are fed into complex neural network architectures that automatically determine the important acoustic information and distinguish between healthy and dysarthric speech [21,22]. Deep learning approaches require less data preparation and feature engineering but may suffer from a lack of interpretability as further post-processing is often required to interpret how the speaker's speech is impaired.…”