Automatic Phonemic Transcription Using Feature Fusing of Articulatory Attribute and Syllable Type

In this paper, we aim to improve the accuracy of automatic phonemic transcription which takes as inputs frame-level feature sequences and outputs phoneme sequences. The more stable and abstract linguistic knowledge, including the syllable type and the articulatory attribute, are explored and developed as inputting feature. We formally employ the designated bottleneck layer of the DNN model to obtain the feature representation from the two information. Then, the two types of features are connected to form the final feature representation. Based on the learned feature representation, the automatic phonemic transcription is processed as the end-to-end sequence labelling, and the encoder-decoder architecture is chosen to dealt with different lengths in both input and output. These experimental results on two languages shown that the learned feature outperforms the traditional acoustic feature, and the proposed method achieves the phoneme accuracy improvements.