Abstract: An identity vector generation method is provided. The method includes obtaining to-be-processed speech data. Corresponding acoustic features are extracted from the to-be-processed speech data. A posterior probability that each of the acoustic features belongs to each Gaussian distribution component in a speaker background model is calculated to obtain a statistic. The statistic is mapped to a statistic space to obtain a reference statistic, the statistic space built according to a statistic corresponding to a speech sample exceeding a threshold speech duration. A corrected statistic is determined according to the calculated statistic and the reference statistic; and an identity vector is generated according to the corrected statistic.

Abstract: A method and an apparatus for training a voiceprint recognition system are provided. The method includes obtaining a voice training data set comprising voice segments of users; determining identity vectors of all the voice segments; identifying identity vectors of voice segments of a same user in the determined identity vectors; placing the recognized identity vectors of the same user in the users into one of user categories; and determining an identity vector in the user category as a first identity vector. The method further includes normalizing the first identity vector by using a normalization matrix, a first value being a sum of similarity degrees between the first identity vector in the corresponding category and other identity vectors in the corresponding category; training the normalization matrix, and outputting a training value of the normalization matrix when the normalization matrix maximizes a sum of first values of all the user categories.

Abstract: A method is performed by at least one processor, and includes acquiring training speech data by concatenating speech segments having a lowest target cost among candidate concatenation solutions, and extracting training speech segments of a first annotation type, from the training speech data, the first annotation type being used for annotating that a speech continuity of a respective one of the training speech segments is superior to a preset condition.

Abstract: The embodiments of the present disclosure disclose a method for constructing a speech decoding network in digital speech recognition. The method comprises acquiring training data obtained by digital speech recording, the training data comprising a plurality of speech segments, and each speech segment comprising a plurality of digital speeches; performing acoustic feature extraction on the training data to obtain a feature sequence corresponding to each speech segment; performing progressive training starting from a mono-phoneme acoustic model to obtain an acoustic model; acquiring a language model, and constructing a speech decoding network by the language model and the acoustic model obtained by training.

Abstract: A voice data processing method and apparatus are provided. The method includes obtaining an I-Vector vector of each of voice samples, and determining a target seed sample in the voice samples. A first cosine distance is calculated between an I-Vector vector of the target seed sample and an I-Vector vector of a target remaining voice sample, where the target remaining voice sample is a voice sample other than the target seed sample in the voice samples. A target voice sample is filtered from the voice samples or the target remaining voice sample according to the first cosine distance, to obtain a target voice sample whose first cosine distance is greater than a first threshold.

Abstract: Processing circuitry of an information processing apparatus obtains a set of identity vectors that are calculated according to voice samples from speakers. The identity vectors are classified into speaker classes respectively corresponding to the speakers. The processing circuitry selects, from the identity vectors, first subsets of interclass neighboring identity vectors respectively corresponding to the identity vectors and second subsets of intraclass neighboring identity vectors respectively corresponding to the identity vectors. The processing circuitry determines an interclass difference based on the first subsets of interclass neighboring identity vectors and the corresponding identity vectors; and determines an intraclass difference based on the second subsets of intraclass neighboring identify vectors and the corresponding identity vectors.

Abstract: A method is performed by at least one processor, and includes acquiring training speech data by concatenating speech segments having a lowest target cost among candidate concatenation solutions, and extracting training speech segments of a first annotation type, from the training speech data, the first annotation type being used for annotating that a speech continuity of a respective one of the training speech segments is superior to a preset condition.

Abstract: A method and an apparatus for training a voiceprint recognition system are provided. The method includes obtaining a voice training data set comprising voice segments of users; determining identity vectors of all the voice segments; identifying identity vectors of voice segments of a same user in the determined identity vectors; placing the recognized identity vectors of the same user in the users into one of user categories; and determining an identity vector in the user category as a first identity vector. The method further includes normalizing the first identity vector by using a normalization matrix, a first value being a sum of similarity degrees between the first identity vector in the corresponding category and other identity vectors in the corresponding category; training the normalization matrix, and outputting a training value of the normalization matrix when the normalization matrix maximizes a sum of first values of all the user categories.

Abstract: An identity vector generation method is provided. The method includes obtaining to-be-processed speech data. Corresponding acoustic features are extracted from the to-be-processed speech data. A posterior probability that each of the acoustic features belongs to each Gaussian distribution component in a speaker background model is calculated to obtain a statistic. The statistic is mapped to a statistic space to obtain a reference statistic, the statistic space built according to a statistic corresponding to a speech sample exceeding a threshold speech duration. A corrected statistic is determined according to the calculated statistic and the reference statistic; and an identity vector is generated according to the corrected statistic.

Abstract: A voice data processing method and apparatus are provided. The method includes obtaining an I-Vector vector of each of voice samples, and determining a target seed sample in the voice samples. A first cosine distance is calculated between an I-Vector vector of the target seed sample and an I-Vector vector of a target remaining voice sample, where the target remaining voice sample is a voice sample other than the target seed sample in the voice samples. A target voice sample is filtered from the voice samples or the target remaining voice sample according to the first cosine distance, to obtain a target voice sample whose first cosine distance is greater than a first threshold.

Abstract: The embodiments of the present disclosure disclose a method for constructing a speech decoding network in digital speech recognition. The method comprises acquiring training data obtained by digital speech recording, the training data comprising a plurality of speech segments, and each speech segment comprising a plurality of digital speeches; performing acoustic feature extraction on the training data to obtain a feature sequence corresponding to each speech segment; performing progressive training starting from a mono-phoneme acoustic model to obtain an acoustic model; acquiring a language model, and constructing a speech decoding network by the language model and the acoustic model obtained by training.

Abstract: Processing circuitry of an information processing apparatus obtains a set of identity vectors that are calculated according to voice samples from speakers. The identity vectors are classified into speaker classes respectively corresponding to the speakers. The processing circuitry selects, from the identity vectors, first subsets of interclass neighboring identity vectors respectively corresponding to the identity vectors and second subsets of intraclass neighboring identity vectors respectively corresponding to the identity vectors. The processing circuitry determines an interclass difference based on the first subsets of interclass neighboring identity vectors and the corresponding identity vectors; and determines an intraclass difference based on the second subsets of intraclass neighboring identify vectors and the corresponding identity vectors.