Apparatus and method for performing model estimation utilizing a discriminant measure

A method for performing acoustic model estimation in order to optimize classification accuracy on feature vectors derived from a speaker with respect to a plurality of classes corresponding to phones to which a plurality of acoustic models respectively correspond, the method comprising the steps of:

• (a) initializing an acoustic model for each phone;
• (b) evaluating the merit of the acoustic model initialized for each phone utilizing an objective function having a two component discriminant measure capable of characterizing each phone whereby a first component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from the phone and a second component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from other phones;
• (c) adapting the acoustic model for selected phones so as to one of increase the first component of the discriminant measure for the phone and decrease the second component of the discriminant measure for the phone, the adapting step yielding a new acoustic model for each selected phone;
• (d) evaluating the merit of the new acoustic models for each phone adapted in step (c) utilizing the two component discriminant measure;
• (e) comparing results of the evaluation performed in step (b) with results of the evaluation of step (d) for each phone, and if one of the first component of the discriminant measure has increased and the second component of the discriminant measure has decreased, then the new acoustic model is kept for that phone, else the acoustic model originally initialized is kept;
• (f) estimating parameters associated with each acoustic model kept for each phone in order to substantially optimize the objective function;
• and
• (g) evaluating termination criterion to determine if the parameters of the acoustic models are substantially optimized.

A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for performing acoustic model estimation in order to optimize classification accuracy on feature vectors derived from a speaker with respect to a plurality of classes corresponding to phones to which a plurality of acoustic models respectively correspond, the method comprising the steps of:

• (a) initializing an acoustic model for each phone;
• (b) evaluating the merit of the acoustic model initialized for each phone utilizing an objective function having a two component discriminant measure capable of characterizing each phone whereby a first component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from the phone and a second component is defined as a probability that the acoustic model for the phone assigns to the feature vectors from other phones;
• (c) adapting the acoustic model for selected phones so as to one of increase the first component of the discriminant measure for the phone and decrease the second component of the discriminant measure for the phone, the adapting step yielding a new acoustic model for each selected phone;
• (d) evaluating the merit of the new acoustic models for each phone adapted in step (c) utilizing the two component discriminant measure;
• (e) comparing results of the evaluation performed in step (b) with results of the evaluation of step (d) for each phone, and if one of the first component of the discriminant measure has increased and the second component of the discriminant measure has decreased, when the new acoustic model is kept for that phone, else the acoustic model originally initialized is kept;
• (f) estimating parameters associated with each acoustic model kept for each phone in order to substantially optimize the objective function;
• and
• (g) evaluating termination criterion to determine if the parameters of the acoustic models are substantially optimized.