# Speech Disorder Classification Using Extended Factorized Hierarchical Variational Auto-encoders

**Authors:** Jinzi Qi, Hugo Van hamme  
**arXiv:** [2106.07337](https://arxiv.org/abs/2106.07337)  
**Date:** 2021-06-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.07337v1/

## Key points

- Research focuses on the classification of objective speech disorders in individuals with communication difficulties
- Neural networks are proposed for this application due to advancements in speech technology
- An extended version of Factorized Hierarchical Variational Auto-encoders (FHVAE) is applied for representation learning on disordered speech
- FHVAE model extracts content-related and sequence-related latent variables from speech data
- Latent variables are aggregated at both word and sentence levels for improved classification performance
- Study demonstrates successful disentanglement of content-related and sequence-related representations, but both are necessary for optimal results in disorder type classification
- COPAS database is utilized, containing recordings of Dutch Intelligibility Assessment (DIA) from 319 speakers with various types of speech disorders
- Only five types of speakers are selected from the database: control speakers, dysarthria, cleft palate, laryngectomy, and impaired speech secondary to hearing impairment
- DIA recordings are segmented into word-level audio pieces based on provided word-level alignment
- Data augmentation techniques such as altering speed rate using 0.8x or 1.2x are employed to increase available sample size for training models
- Overall research highlights successful application of extended FHVAE model for representation learning in classification of speech disorders, emphasizing importance of both content-related and sequence-related representations

## Summary

This research focuses on the classification of objective speech disorders in individuals with communication difficulties. The use of neural networks for this application is proposed due to advancements in speech technology. To address the lack of labeled disordered speech data, an extended version of Factorized Hierarchical Variational Auto-encoders (FHVAE) is applied for representation learning on disordered speech. The FHVAE model extracts both content-related and sequence-related latent variables from speech data which are then utilized to explore how information about the type of disorder is represented in the latent variables. For improved classification performance, the latent variables are aggregated at both the word and sentence levels. The study demonstrates that the extension of the FHVAE model successfully disentangles content-related and sequence-related representations; however, it is found that both representations are still necessary for achieving optimal results in disorder type classification. The COPAS database is utilized which contains recordings of Dutch Intelligibility Assessment (DIA) along with other materials from 319 speakers with various types of speech disorders such as dysarthria, voice disorders, cleft palate, articulation disorders, laryngectomy, glossectomy and impaired speech secondary to hearing impairment. To ensure a balanced dataset with sufficient data sizes only five types of speakers (control speakers, dysarthria, cleft palate, laryngectomy and impaired speech secondary to hearing impairment) are selected from the database consisting each type consists 29 speakers. The DIA recordings are segmented into word-level audio pieces based on provided word-level alignment and due to limited data availability (approximately 1.6 hours), data augmentation techniques such as altering speed rate using 0.8x or 1.2x are employed to increase available samples size for training models. Overall this research highlights successful application of an extended FHVAE model for representation learning in classification of speech disorders emphasizing importance of both content-related and sequence-related representations in achieving accurate disorder type classification results.