Data-driven deep learning solutions, which are gradient-based neural architectures, have proven useful in overcoming some limitations of traditional signal processing techniques. However, a large number of reverberated-anechoic training utterance pairs covering as many environmental conditions as possible is required to achieve robust performance in unseen testing conditions. In this study, we propose to address the data requirement issue while preserving the advantages of deep neural structures leveraging upon hierarchical extreme learning machines (HELMs), which are not gradient-based neural architectures. In particular, an ensemble HELM learning framework is established to effectively recover anechoic speech from a reverberated one based on a spectral mapping. In addition to the ensemble learning framework, we further derive two novel HELM models, namely highway HELM, termed HELM(Hwy), and residual HELM, termed HELM(Res), both incorporating low-level features to enrich the information for spectral mapping. We evaluated the proposed ensemble learning framework using simulated and measured impulse responses by employing TIMIT, MHINT, and REVERB corpora. Experimental results show that the proposed framework outperforms both traditional methods and a recently proposed integrated deep and ensemble learning algorithm in terms of standardized objective and subjective evaluations under matched and mismatched testing conditions for simulated and measured impulse responses.
|期刊||IEEE Transactions on Cognitive and Developmental Systems|
|出版狀態||接受/付印 - 2019 一月 1|
ASJC Scopus subject areas
- Artificial Intelligence