Framewise phoneme classification with bidirectional LSTM and other neural network architectures

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Recurrent Neural Networks (RNNs) and Multilayer Perceptrons (MLPs) are standard architectures for speech processing.
• Exploiting contextual information is crucial for accurate speech recognition.

Purpose of the Study:

• To introduce bidirectional Long Short Term Memory (BLSTM) networks and a modified LSTM learning algorithm.
• To evaluate the performance of BLSTM against other architectures for framewise phoneme classification.

Main Methods:

• Implementation of bidirectional Long Short Term Memory (LSTM) networks.
• Utilizing a modified, full gradient LSTM learning algorithm.
• Benchmarking on the TIMIT database for framewise phoneme classification.

Main Results:

• Bidirectional networks demonstrated superior performance compared to unidirectional networks.
• Long Short Term Memory (LSTM) networks were found to be faster and more accurate than standard RNNs and time-windowed MLPs.
• BLSTM networks proved effective in leveraging contextual information for speech processing.

Conclusions:

• Bidirectional networks are more effective than unidirectional ones for speech processing tasks.
• LSTM networks, particularly BLSTM, offer significant advantages in speed and accuracy.
• BLSTM is a highly effective architecture for exploiting contextual information in speech processing.