Imed Laaridh

Automatic Detection of Phone-Based Anomalies in Dysarthric Speech

Perceptual evaluation is still the most common method in clinical practice for diagnosing and following the progression of the condition of people with speech disorders. Although a number of studies have addressed the acoustic analysis of speech productions exhibiting impairments, additional descriptive analysis is required to manage interperson variability, considering speakers with the same condition or across different conditions. In this context, this article investigates automatic speech processing approaches dedicated to the detection and localization of abnormal acoustic phenomena in speech signal produced by people with speech disorders. This automatic process aims at enhancing the manual investigation of human experts while at the same time reducing the extent of their intervention by calling their attention to specific parts of the speech considered as atypical from an acoustical point of view. Two different approaches are proposed in this article. The first approach models only the normal speech, whereas the second models both normal and dysarthric speech. Both approaches are evaluated following two strategies: one consists of a strict phone comparison between a human annotation of abnormal phones and the automatic output, while the other uses a “one-phone delay” for the comparison. The experimental evaluation of both approaches for the task of detecting acoustic anomalies was conducted on two different corpora composed of French dysarthric speakers and control speakers. These approaches obtain very encouraging results and their potential for clinical uses with different types of dysarthria and neurological diseases is quite promising.

Perceptual Evaluation for Automatic Anomaly Detection in Disordered Speech : focus on ambiguous cases

Abstract Perceptual evaluation is still the most common method in clinical practice for diagnosing and following the condition progression of people suffering from dysarthria (or speech disorders more generally). Such evaluations are frequently described as non-trivial, subjective and highly time-consuming (depending on the evaluation level). Most of the time, perceptual assessment is performed individually by clinicians which can be problematic since judgment may vary from one clinician to the other. Clinicians have therefore expressed the need for new objective evaluation tools better adapted to longitudinal studies, the observation of small units and rehabilitation context to monitor patients’ progress. We have previously proposed an automatic approach to the anomaly detection at the phone level for dysarthric speech. The system behavior was studied and validated with different corpora and speech styles and shows good results in this specific task. Nonetheless, the lack of annotated French dysarthric speech corpora has limited our ability to analyze some aspects of its behavior, such as severity, more precisely (more anomalies are detected automatically compared with human experts). To overcome this limitation, we proposed an original perceptual evaluation protocol applied to a limited set of decisions made by the automatic system, relating to the presence of anomalies. Particularly, we intended to focus our analyses on some ambiguous cases in order to enrich our knowledge about the system behavior. This evaluation was carried out by a jury of 29 non-naive individuals. Results confirm the relevance of the system for the anomaly detection, and place it within the most severe juries. Besides interesting information related to the system behavior, the evaluation protocol highlighted main differences between human process and the automatic system: humans have difficulties in focusing on small units and they are influenced by contextual information, while the system only focuses on small units. In a way, both approaches are probably complementary.

Evaluation of a Phone-Based Anomaly Detection Approach for Dysarthric Speech.

Perceptual evaluation is still the most common method in clinical practice for the diagnosing and the following of the condition progression of people with speech disorders. Many automatic approaches were proposed to provide objective tools to deal with speech disorders and help professionals in the severity evaluation of speech impairments. This paper investigates an automatic phone-based anomaly detection approach implying an automatic text-constrained phone alignment. Here, anomalies are related to speech segments, for which an unexpected acoustic pattern is observed, compared with a normal speech production. This objective tool is applied to French dysarthric speech recordings produced by patients suffering from four different pathologies. The behavior of the anomaly detection approach is studied according to the precision of the automatic phone alignment. Faced with the difficulties of having a gold standard reference, especially for the phone-based anomaly annotation, this behavior is observed on both annotated and nonannotated corpora. As expected, alignment errors (large shifts compared with a manual segmentation) lead to a large amount of anomalies automatically detected. However, about 50% of correctly detected anomalies are not related to alignment errors. This behavior shows that the automatic approach is able to catch irregular acoustic patterns of phones.