	US 7,552,049 B2
	Noise adaptation system of speech model, noise adaptation method, and noise adaptation program for speech recognition
Zhipeng Zhang, Tokyo (Japan); Kiyotaka Otsuji, Kamakura (Japan); Toshiaki Sugimura, Yokohama (Japan); and Sadaoki Furui, Tokyo (Japan)
Assigned to NTT DoCoMo, Inc., Tokyo (Japan); and Sadaoki Furui, Tokyo (Japan)
Filed on Mar. 10, 2004, as Appl. No. 10/796,283.
Claims priority of application No. 2003-066933 (JP), filed on Mar. 12, 2003.
Prior Publication US 2004/0204937 A1, Oct. 14, 2004
This patent is subject to a terminal disclaimer.
Int. Cl. G10L 15/00 (2006.01)

U.S. Cl. 704—231 [704/233; 704/245; 704/250]

8 Claims

1. A noise adaptation system of speech model for adapting a speech model for any noise to speech to be recognized in a noisy environment, said speech model being learned by using clean speech data, said system comprising:

clustering means for clustering noise-added speech;

speech model space generating means for generating a tree-structure noisy speech model space based on the result of the clustering performed by said clustering means;

parameter extracting means for extracting a speech feature parameter of input noisy speech to be recognized;

selecting means for selecting an optimum model from the tree-structure noisy speech model space generated by said speech model space generating means; and

linear transformation means for applying linear transformation to the model selected by the selecting means so that the model provides a further increased likelihood,

wherein said clustering means generates said noise-added speech by adding said noise to said speech in accordance with a signal-to-noise ratio condition, subtracts the mean value of speech cepstral of the generated noise-added speech, generates a Gaussian distribution model of each of pieces of generated noise-added speech, and calculates the likelihood between the pieces of noise-added speech to generate a likelihood matrix to provide a clustering result.