Speech recognition method, program and apparatus using multiple acoustic models

A speech recognition method, wherein speech data on which different types of noise have been superposed respectively are created, the noise is eliminated by a spectral subtraction method or a continuous spectral subtraction method from each of the speech data on which the different types of noise have been superposed, a cepstrum mean normalization method is applied to each of the speech data which have undergone the noise elimination to obtain feature vectors of a speech segment, and acoustic models corresponding to each of the noise types are created and stored based on the feature vectors;

• when a speech recognition is performed, a first speech feature analysis is performed to obtain frequency-domain feature data of speech data to be recognized;
• a determination is made whether the speech data is a noise segment or a speech segment based on the result of the feature analysis, and when a noise segment is detected, the feature vector thereof is stored, and when a speech segment is detected, the noise is eliminated from the speech segment by the spectral subtraction method or the continuous spectral subtraction method;
• a second speech feature analysis is performed on the speech segment data which has undergone the noise elimination to obtain cepstrum coefficients, and the feature vector of the speech segment is stored;
• when the speech segment has terminated, the type of the noise superposed is determined based on the feature data of the noise segment having been stored, and an acoustic model is selected from the acoustic models corresponding to each of the noise types;
• the cepstrum mean normalization method is applied to the feature vector of the speech segment on which the noise has been superposed, using the feature vector of the speech segment having been stored, to obtain the feature vector of the speech segment;
• and
• a speech recognition is performed on the feature vector obtained by the cepstrum mean normalization method based on the selected acoustic model.

A speech recognition method, comprising:

• creating speech data on which a particular type of noise with different S/N ratios have been superposed respectively;
• eliminating the noise by a predetermined noise elimination method from each of the speech data;
• creating and storing acoustic models corresponding to each of the S/N ratios using feature vectors of each of the speech data which have undergone the noise elimination;
• when a speech recognition is performed, determining the S/N ratio of a noise superposed on speech data to be recognized;
• selecting a corresponding acoustic model from the acoustic models corresponding to each of the S/N ratios based on the result of the determination;
• eliminating the noise by the predetermined noise elimination method from the speech data to be recognized on which the noise has been superposed;
• and
• performing a speech recognition on the feature vector of the speech data which has undergone the noise elimination based on the selected acoustic model,
• wherein, when speech data on which another type of noise also with different S/N ratios have been superposed is created, other acoustic models are created corresponding to each of the S/N ratios of the other noise type.

A storage medium storing a speech recognition program, the speech recognition program, the speech recognition program comprising the steps of:

• creating speech data on which different types of noise have been superposed respectively, eliminating the noise by a predetermined noise elimination method from each of the speech data on which the noise has been superposed, and creating acoustic models corresponding to each of a plurality of S/N ratios of each of the noise types using the feature vectors obtained by analyzing the features of each of the speech data which have undergone the noise elimination, and storing the acoustic models in acoustic model storage device;
• determining the type of a noise superposed on speech data to be recognized, and selecting a corresponding acoustic model from the acoustic models stored in said acoustic model storage device based on the result of the determination;
• eliminating the noise by the predetermined noise elimination method from the speech data to be recognized on which the noise has been superposed;
• and
• performing a speech recognition on the feature vector of the speech data which has undergone the noise elimination based on the selected acoustic model.

A storage medium storing a speech recognition program, the speech recognition program comprising the steps of:

• creating speech data on which different types of noise have been superposed respectively, eliminating the noise by at least one of a spectral subtraction method and a continuous spectral subtraction method from each of the speech data on which the different types of noise have been superposed, applying a cepstrum mean normalization method to each of the speech data which have undergone the noise elimination to obtain the feature vectors of a speech segment, and creating acoustic models corresponding to each of a plurality of S/N ratios of each of the noise types based on the feature vectors and storing the acoustic models in acoustic model storage means;
• performing a first speech feature analysis to obtain frequency-domain feature data of speech data to be recognized on which a noise has been superposed;
• determining whether the speech data is a noise segment or a speech segment based on the result of the feature analysis, and storing the feature vector thereof when a noise segment is detected;
• eliminating the noise from the speech segment by the at least one of the spectral subtraction method and the continuous spectral subtraction method when a speech segment is detected;
• performing a second speech feature analysis on the speech segment data which has undergone the noise elimination to obtain cepstrum coefficients, and storing the feature vector of the speech segment;
• when the speech segment has terminated, determining the type of the noise superposed based on the feature data of the noise segment having been stored, and selecting an acoustic model from the acoustic models corresponding to each of the noise types;
• applying the cepstrum mean normalization method to the feature vector of the speech segment on which the noise has been superposed, using the feature vector of the speech segment having been stored, to obtain the feature vector of the speech segment;
• and
• performing a speech recognition on the feature vector obtained by the cepstrum mean normalization method based on the selected acoustic model.

A storage medium storing a speech recognition program, the speech recognition program comprising the steps of:

• creating speech data on which a particular type of noise with different S/N ratios have been superposed respectively, eliminating the noise by a predetermined noise elimination method from each of the speech data, and creating acoustic models corresponding to each of the S/N ratios using the feature vectors of each of the speech data which have undergone the noise elimination and storing the acoustic models in acoustic model storage device;
• determining the S/N ratio of a noise superposed on speech data to be recognized, and selecting a corresponding acoustic model from the acoustic models corresponding to each of the S/N ratios based on the result of the determination;
• eliminating the noise by the predetermined noise elimination method from the speech data to be recognized on which the noise has been superposed;
• and
• performing a speech recognition on the feature vector of the speech data which has undergone the noise elimination based on the selected acoustic model,
• wherein, when speech data on which another type of noise also with different S/N ratios have been superposed in created, other acoustic models are created corresponding to each of the S/N ratios of the other noise type.

A speech recognition apparatus, comprising:

• acoustic models corresponding to each of a plurality of S/N ratios of each of different types of noise, created by generating speech data on which the different types of noise have been superposed respectively, eliminating the noise by a predetermined noise elimination method from each of the speech data on which the different types of noise have been superposed, and using the feature vectors of each of the speech data which have undergone the noise elimination;
• an acoustic model storage device that stores the acoustic models;
• a noise determination device that determines the type of a noise superposed on speech data to be recognized;
• an acoustic model selection device that selects a corresponding acoustic model from the acoustic models corresponding to each of the noise types based on the result of the determination;
• a noise elimination device that eliminates the noise by the predetermined noise elimination method from the speech data to be recognized on which the noise has been superposed;
• and
• a speech recognition device that performs a speech recognition on the feature vector of the speech data which has undergone the noise elimination based on the selected acoustic model.

A speech recognition apparatus, comprising:

• acoustic models corresponding to each of a plurality of S/N ratios of each of different types of noise, created by generating speech data on which the different types of noise have been superposed respectively, eliminating the noise by a spectral subtraction method or a continuous spectral subtraction method from each of the speech data on which the different types of noise have been superposed, applying a cepstrum mean normalization method to each of the speech data which have undergone the noise elimination to obtain the feature vectors of a speech segment, and using the feature vectors;
• an acoustic model storage device that stores the acoustic models;
• a first speech feature analysis device that performs a first speech feature analysis to obtain frequency-domain feature data of speech data to be recognized;
• a noise segment/speech segment determination device that determines whether the speech data is a noise segment or a speech segment based on the result of the feature analysis, and stores the feature vector thereof in feature data storage means when a noise segment is detected;
• a noise elimination device that eliminates the noise from the speech segment by the spectral subtraction method or the continuous spectral subtraction method when a speech segment is detected;
• a second speech feature analysis device that performs a second speech feature analysis on the speech segment data which has undergone the noise elimination to obtain cepstrum coefficients, and storing the feature vector of the speech segment in the feature data storage device;
• a noise type determination device that determines, when the speech segment has terminated, the type of the noise superposed based on the feature data of the noise segment having been stored;
• an acoustic model selection device that selects a corresponding acoustic model from the acoustic models corresponding to each of the noise types;
• a cepstrum mean normalization operation device that applies the cepstrum mean normalization method to the feature vector of the speech segment on which the noise has been superposed, using the feature vector of the speech segment having been stored, to output the feature vector of the speech segment;
• and
• a speech recognition device that performs a speech recognition on the feature vector based on the selected acoustic model.

A speech recognition apparatus, comprising:

• acoustic models corresponding to each of different S/N ratios for a particular type of noise, created by generating speech data on which the particular type of noise with the different S/N ratios have been superposed respectively, eliminating the noise by a predetermined noise elimination method from each of the speech data, and using the feature vectors of each of the speech data which have undergone the noise elimination;
• an acoustic models storage device that stores the acoustic models;
• a S/N ratio determination means for determining the S/N ratio of a noise superposed on speech data to be recognized;
• an acoustic model selection device that selects a corresponding acoustic model from the acoustic models corresponding to each of the S/N ratios based on the result of the determination;
• a noise elimination device that eliminates the noise by the predetermined noise elimination method from the speech data to be recognized on which the noise has been superposed;
• and
• a speech recognition device that performs a speech recognition on the feature vector of the speech data which has undergone the noise elimination based on the selected acoustic model,
• wherein, when speech data on which another type of noise also with different S/N ratios have been superposed is created, other acoustic models are created corresponding to each of the S/N ratios of the other noise type.