Updating hidden conditional random field model parameters after processing individual training samples

#In phonetic classification and speech recognition, Hidden Markov Models (HMMs) have been used extensively to model the acoustics of speech. HMMs are generative models that use the concept of a hidden state sequence to model the non-stationary of the generation of observations from a label. At each frame of an input signal, the HMM determines the probability of generating that frame from each possible hidden state. This probability is determined by applying a feature vector derived from the frame of speech to a set of probability distributions associated with the state. In addition, the HMM determines a probability of transitioning from a previous state to each of the states in the Hidden Markov Model. Using the combined transition probability and observation probability, the Hidden Markov Model selects a state that is most likely to have generated the frame.

#One limitation of Hidden Markov Models is that the probabilities of each state are determined using the same feature vectors, and thus the same collection of features are used against each state. This limitation is undesirable because different features are more important for certain speech sounds than for others. For example, when distinguishing vowel sounds from each other, the value of the formants are important. However, when distinguishing between fricatives, information as to whether the speech is voiced or unvoiced is informative. However, HMM systems do not allow the system to be able to use different features for states associated with different speech sounds.

#In addition, HMM models do not allow a change in the length of between-frame dependencies for the features. Thus, at each frame, the features provide a fixed amount of information about previous frames. To help distinguish between speech sounds, current systems do not allow for different length frame dependencies for states associated with different speech sounds. Also, HMM models do not allow any flexibility in the amount of acoustic data that is summarized in each frame. Typically, frames are generated by analyzing 25 millisecond segments of acoustic waveform. This is a compromise between the long time scales required for frequency analysis of voiced sounds such as vowels and the short time scales required for reliably detecting short sounds such as plosives.

#FIG. 1 is a block diagram of one computing environment in which the present invention may be practiced.

#FIG. 2 is a block diagram of a speech recognition/phonetic classifier training and decoding system under one embodiment of the present invention.

#FIG. 3 is a flow diagram for training a hidden conditional random field model in accordance with one embodiment of the present invention.

#The invention relates to training parameters in hidden conditional random field models. However, before describing the invention in more detail, one illustrative environment in which the present invention can be used will be described.

#FIG. 1 illustrates an example of a suitable computing system environment 100 on which the invention may be implemented. The computing system environment 100 is only one example of a suitable computing environment and is not intended to suggest any limitation as to the scope of use or functionality of the invention. Neither should the computing environment 100 be interpreted as having any dependency or requirement relating to any one or combination of components illustrated in the exemplary operating environment 100.

#The invention is operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well-known computing systems, environments, and/or configurations that may be suitable for use with the invention include, but are not limited to, personal computers, server computers, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, telephony systems, distributed computing environments that include any of the above systems or devices, and the like.