Transfer Learning for Tandem ASR Feature Extraction
Tandem automatic speech recognition (ASR), in which one or an ensemble of multi-layer perceptrons (MLPs) is used to provide a non-linear transform of the acoustic parameters, has become a standard technique in a number of state-of-the-art systems. In this paper, we examine the question of how to transfer learning from out-of-domain data to new tasks.
Our primary focus is to develop tandem features for recognition of speech from the meetings domain. We show that adapting MLPs originally trained on conversational telephone speech leads to lower word error rates than training MLPs solely on the target data. Multi-task learning, in which a single MLP is trained to perform a secondary task (in this case a speech enhancement mapping from farfield to nearfield signals) is also shown to be advantageous.
We also present recognition experiments on broadcast news data which suggest that structure learned from English speech can be adapted to Mandarin Chinese. The performance of tandem MLPs trained on 440 hours of Mandarin speech with a random initialization was achieved by adapted MLPs using about 97 hours of data in the target language.
Unable to display preview. Download preview PDF.