Progressive AutoSpeech: An Efficient and General Framework for Automatic Speech Classification

摘要

Speech classification has been widely used in many speech-related applications. However, the complexity of speech classification tasks often exceeds the scope of non-experts, the off-the-shelf speech classification methods are urgently needed. Recently, the automatic speech classification (AutoSpeech) without any human intervention has attracted more and more attention. The practical AutoSpeech solution should be general and can automatically handle classification tasks from different domains. Moreover, AutoSpeech should improve not only the final performance but also the any-time performance especially when the time budget is limited. To address these issues, we propose a three-stage any-time learning algorithm framework called Progressive AutoSpeech for automatic speech classification under a given time budget. Progressive AutoSpeech consists of the fast stage, enhancement stage, and exploration stage. Each stage uses different models and features to ensure generalization. Additionally, we automatically construct ensembles of top-k; prediction results to improve the robustness. The experimental results reveal that Progressive AutoSpeech is effective and efficient for a wide range of speech classification tasks and can achieve the best ALC score.