Transfer of an acoustic knowledge to a neural network

A system comprising:

• a processor;
• and
• a memory storing a program, which, when executed on the processor, performs a method for transferring acoustic knowledge of a trained acoustic model (AM) to a neural network (NN), the method comprising:
• reading, into the memory, the NN and the AM, the AM being trained with target domain data, and a set of training data including a set of frames wherein each frame corresponds to one phoneme data, the set of training data being data obtained from a domain different from a target domain for the target domain data;
• inputting training data from the set of training data into the AM, calculating one or more posterior probabilities of context-dependent states corresponding to each phoneme in a phoneme class for the phoneme data associated with each frame in the training data, and generating a posterior probability vector from the one or more posterior probabilities as a soft label for the NN, each element of the soft label for the NN being limited to a posterior probability, from the one or more posterior probabilities, associated with each respective phoneme model symbol corresponding to the input phoneme data;
• replacing one or more of the one or more posterior probabilities of the soft label for the NN determined to be less than a predetermined threshold with a constant value, the constant value being smaller than the replaced posterior probabilities;
• and
• inputting the training data into the NN and updating the NN, using the soft label for the NN.

A non-transitory computer program product for transferring acoustic knowledge of a trained acoustic model (AM) to a neural network (NN), the computer program product comprising a computer readable storage medium having program instructions embodied therewith, wherein the computer readable storage medium is not a transitory signal per se, the program instructions executable by a computer to cause the computer to perform a method comprising:

• reading, into memory, the NN and the AM, the AM being trained with target domain data, and a set of training data including a set of frames wherein each frame corresponds to one phoneme data, the set of training data being data obtained from a domain different from a target domain for the target domain data;
• inputting training data from the set of training data into the AM, calculating one or more posterior probabilities of context-dependent states corresponding to each phoneme in a phoneme class for the phoneme data associated with each frame in the training data, and generating a posterior probability vector from the one or more posterior probabilities as a soft label for the NN, each element of the soft label for the NN being limited to a posterior probability, from the one or more posterior probabilities, associated with each respective phoneme model symbol corresponding to the input phoneme data;
• replacing one or more of the one or more posterior probabilities of the soft label for the NN determined to be less than a predetermined threshold with a constant value, the constant value being smaller than the replaced posterior probabilities;
• and
• inputting the training data into the NN and updating the NN, using the soft label for the NN.