This repo contains customized Kaldi recipes intended for training DNN feature extractors. That is, for extracting meaningful feature embeddings from the utterances which can be later utilized for ML classification tasks. Each recipe has its own dataset. Some of the datasets are open-sourced, while others like the Dementia or Alzheimer's corpora are not. They can be requested for use though. Just drop me a message. This repo also contains a recipe for doing word and phone level forced-alignments using the Tedlium-v3 corpus.
The corpora used is related to both medical speech and computational paralinguistics:
Sleepiness Corpus -> [B.W. Schuller et al., “The INTERSPEECH2019 Computational Paralinguistics Challenge: StyrianDialects, Continuous Sleepiness, Baby Sounds & OrcaActivity.,”inProceedings of Interspeech, 2019, pp.2378–2382.]
Mask --> [B. W. Schuller et al., “The INTERSPEECH 2020 Computational Paralinguistics Challenge: Elderlyemotion, Breathing & Masks,” inProceedings of Interspeech,Shanghai, China, September 2020]
Emotion Aibo --> [Batliner, Anton & Steidl, Stefan & Noeth, Elmar. (2008). Releasing a thoroughly annotated and processed spontaneous emotional database: the FAU Aibo Emotion Corpus.]
Dementia94B and DementiaNew8k --> Recorded at the Memory Clinic, epartment of Psychiatry of the University of Szeged, Hungary.
BEA Corpus --> [Neuberger, D. et al., “Development of a large spontaneous speechdatabase of agglutinative Hungarian language,” inProceedings ofTSD, Brno, Czech Republic, Sep 2014, pp. 424–431.]