Last week, I submitted a new pre-print to Arxiv. The link is below.
This paper discusses the importance of accommodating uncertainty in addressing the issue of different models in speech classification. We used several variants of uncertainty-based ensemble learning (aka late fusions) to combine predictions of machine learning methods.
Every machine learning models has uncertainty, which, for instance, can be calculated from logits or probabilities. Entropy can be calculated from distribution. This distribution is multiplication of logits with softmax. Then, we can use several models to find the lowest uncertainty score and infer the label from that score. Or, we can weigh the probabilities by those uncertainty scores. We evaluated four variants of uncertainty-based ensemble learning.
Our work is highly beneficial to the community as it will help you improve the prediction of speech classification using ensemble learning.
Some areas that still need further development are stability and generalization. We see the trends that ensemble learning tends to improve the recognition rate of different models, but it is not always the case. We invite readers to collaborate in addressing the various unanswered questions.
We extend our thanks to AIST for their full support of our research.
Happy reading. We welcome your feedback.
Additional information:
Toolkit used in the research: Nkululeko https://github.com/felixbur/nkululeko
Configuration files: https://github.com/bagustris/nkululeko_ensemble_speech_classficcation
---
This post is based on template from [1].
Reference:
[1] https://medium.com/open-science-indonesia/template-memajang-makalah-di-medsos-190c228dd86a
