Automatic Pitch Accent Detection Using Long Short-Term Memory Neural Networks

Prosody detection is gaining increasingly popularity in the domain of prosody research because of its significance in Text to Sound, Computer-aided pronunciation training (CAPT), etc. Pitch accent is an important part of prosody and many recognition models of both static and dynamic have been investigated for automatic labeling it. Recently, artificial neural networks, especially Recurrent Neural Networks (RNNs) have been applied in pitch accent detection. However, traditional recurrent neural networks are unable to learn and remember over long sequences due to the issue of back-propagated error decay. To solve this problem, this paper investigates the use of Long Short-Term Memory (LSTM) neural networks for automatic pitch accent detection. This paper encodes lexical and syntactic features as binary variables and uses syllable-based acoustic features including syllable duration, syllable energy, features related to the fundamental frequency. Our experimental results show that LSTM-RNNs for pitch accent detection achieves an accuracy of 89.0%, which is better than the results of using classical detection methods by about 83.2%.