Temporal pattern recognition method and apparatus utilizing segment and frame-based models

A method of speech recognition comprising:

• determining a frame-based probability for a sequence of input feature vectors by determining the correspondence between each individual feature vector in the sequence of feature vectors and a respective individual frame state from a sequence of frame states;
• dividing the sequence of feature vectors into sub-sequences of feature vectors;
• determining a segment-based probability for the sequence of input feature vectors by determining the correspondence between the feature vectors in each sub-sequence of feature vectors and a respective segment state from a sequence of segment states;
• and
• combining at least the frame-based probability and the segment-based probability to identify a most likely sequence of patterns that can be represented by the input sequence of feature vectors.

A method of training a pattern recognition system using feature vectors generated from a training signal, the method comprising:

• selecting a unified-framework segmentation for a sequence of feature vectors based on a probability associated with the segmentation, the probability comprising a frame-based probability produced by a frame-based model and a segment-based probability produced by a segment-based model;
• retraining the frame-based model using the segmentation of the sequence of feature vectors and a training pattern sequence;
• and
• retraining the segment-based model using the segmentation of the sequence of feature vectors and the training pattern sequence.

A computer-readable medium having computer-executable components for performing steps comprising:

• determining a frame-based probability for a sequence of input feature vectors by separately determining the likelihood of each feature vector based on a corresponding frame state from a sequence of frame states;
• segmenting the feature vectors into segments;
• determining a segment-based probability for the sequence of input feature vectors by determining the total likelihood of the feature vectors assigned to a segment given a corresponding segment state from a sequence of segment states;
• and
• combining at least the frame-based probability and the segment-based probability to identify a most likely sequence of patterns that can be represented by the input sequence of feature vectors.