Statistical-acoustic-model adaptation method, acoustic-model learning method suitable for statistical-acoustic-model adaptation, storage medium in which parameters for building deep neural network are stored, and computer program for adapting statistical acoustic model

Abstract

The invention provides a statistical-acoustic-model adaptation method with which learning data having specific conditions can be used to efficiently adapt acoustic models using deep neural networks (DNN), and with which accuracy can also be improved. A speaker adaptation method for acoustic models using DNNs includes: a step in which utterance data (90-98) for different speakers is separately stored in first storage devices; a step in which hidden layer modules (112-120) for separate speakers are prepared; a step in which preliminary learning for all the layers (42, 44, 110, 48, 50, 52, 54) in a DNN (80) is performed while switching and selecting the utterance data (90-98), and dynamically substituting a specific layer (110) with the hidden layer modules (112-120) corresponding to the selected utterance data; a step in which the DNN specific layer (110), after the preliminary learning has been completed therefor, is substituted with an initial hidden layer; and a step in which parameters for layers other than the initial hidden layer are fixed, and speech data of a specific speaker is used to perform DNN learning.

Description

technical field [0001] The present invention relates to a deep neural network (hereinafter, abbreviated as "DNN" for simplification of description) used in recognition technologies such as voice recognition, and particularly relates to a technique for improving the learning efficiency of a DNN for specific objects. Background technique [0002] As a method of machine learning, DNN has attracted much attention. DNN is suitable for, for example, image recognition and voice recognition, and it is reported that the relative error rate is reduced by 20 to 30% compared to the conventional one, and can exhibit excellent performance (Non-Patent Documents 1 to 3). [0003] The so-called DNN refers to a neural network with more layers than before. Specifically, DNN includes: an input layer, an output layer, and multiple hidden layers arranged between the input layer and the output layer. The input layer has a plurality of input nodes (neurons: neurons). The output layer has only th...

AI Technical Summary

Problems solved by technology

However, learning can still take a long time when large amounts of data are used for learning

Images