Loading patent…
AssigneeNynex Science & Technology, Inc.
InventorKuansan Wang
Estimating the term…
Methods and apparatus for producing efficiently sized models suitable for pattern recognition purposes are described. Various embodiments are directed to the automated generation, evaluation, and selection of reduced size models from an initial model having a relatively large number of components, e.g., more components than can be stored for a particular intended application. To achieve model size reduction in an automated iterative manner, expectation maximization (EM) model training techniques are combined, in accordance with the present invention, with model size constraints. In one embodiment, a new reduced size model is generated using a LaGrange multiplier from an input model and input size constraints during each iteration of the size reducing model training process. The reduced size model generated during one iteration of the process serves as the input to the next iteration. Scoring, e.g., maximum likelihood scoring, and evaluation steps, in conjunction with stop criteria, are used to determine the number of model size reducing iterations performed, and which reduced size model is selected as the output of the size reduction model training process.
A model training apparatus, comprising:
A method of generating a reduced size speech recognition model, the method comprising the steps of:
Dependencies are inferred from the claim text.
Drawings are temporarily unavailable. Open the PDF at USPTO
No related US applications are recorded in the indexed filing history.
Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice: training
Non-hierarchical techniques: using statistics or function optimisation, e.g. modelling of probability density functions
Hidden Markov Models [HMMs]: training of HMMs
Speech analysis techniques or speech synthesis
Term ended 30 Jul 2017 (estimated)
USPTO file history isn’t available for this patent, so only indexed dates are shown.
The present invention relates to pattern recognition and, more particularly, to methods and apparatus for generating and training models for use in pattern recognition systems, e.g., speech and image recognition systems.
Pattern recognition is a problem which is encountered in many applications today. Two common areas in which pattern recognition problems are found are image and speech processing. In image processing, pattern recognition is used, e.g., to recognize a received image and thereby distinguish it from other possible images. In the speech processing context, pattern recognition techniques are used to, e.g., recognize speech. Numerous other pattern recognition applications exist. Pattern recognition generally involves the use of a statistical model. As will be discussed below, there is a need for improved methods and apparatus for generating and training such statistical models.
In as much as speech recognition is a good example of an application where model generation and training is particularly pertinent, the present invention will be discussed primarily in the context of speech recognition systems. However, the following discussion of the present invention as it applies to speech recognition is intended to facilitate the explanation of the present invention without necessarily limiting the invention to the field of speech recognition. In fact, the methods and apparatus of the present invention may be applied to a wide variety of pattern recognition applications.
FIG. 1 illustrates a known expectation maximization (EM) model training circuit.
FIG. 2 illustrates a known minimum classification error criterion (MCE) model training circuit.
FIG. 3 illustrates a computer system suitable for performing speech recognition model training and speech recognition operations in accordance with the present invention.
As discussed above, the present invention is directed to methods and apparatus for generating improved models for use in pattern recognition systems, e.g., speech and image recognition systems.
FIG. 3 illustrates a computer system 300, suitable for use in performing various model training methods of the present invention. As illustrated, the computer system 300 includes a computer 302, a keyboard/pointing device 304, and a display 306. The computer system 300, is coupled to an external source of acoustic signal and/or training data such as a network data base or other external data source.
The computer 302 includes an input/output (I/O) interface 308 for electrically interfacing the internal components of the computer 302 to the external source of acoustic signals/training data and the external computer system components including the input and output devices, e.g., the keyboard/mouse 304 and the display device 306.