A voice conversion device is provided with a linguistic information extraction unit that extracts linguistic information corresponding to utterance content from a conversion source voice signal, an appearance feature extraction unit that extracts appearance features expressing features related to the look of a person's face from a captured image of the person, and a converted voice generation unit that generates a converted voice on a basis of the linguistic information and the appearance features.

Audio translation system includes a feature extractor and a style transfer machine learning model. The feature extractor generates for each of a plurality of source voice files one or more source voice parameters encoded as a collection of source feature vectors, and generates for each of a plurality of target voice files one or more target voice parameters encoded as a collection of target feature vectors. The style transfer machine learning model trained on the collection of source feature vectors for the plurality of source voice files and the collection of target feature vectors for the plurality of target voice files to generate a style transformed feature vector.

This specification describes a computer-implemented method of generating speech audio for use in a video game, wherein the speech audio is generated using a voice convertor that has been trained to convert audio data for a source speaker into audio data for a target speaker. The method comprises receiving: (i) source speech audio, and (ii) a target speaker identifier. The source speech audio comprises speech content in the voice of a source speaker. Source acoustic features are determined for the source speech audio. A target speaker embedding associated with the target speaker identifier is generated as output of a speaker encoder of the voice convertor. The target speaker embedding and the source acoustic features are inputted into an acoustic feature encoder of the voice convertor. One or more acoustic feature encodings are generated as output of the acoustic feature encoder. The one or more acoustic feature encodings are derived from the target speaker embedding and the source acoustic features. Target speech audio is generated for the target speaker. The target speech audio comprises the speech content in the voice of the target speaker. The generating comprises decoding the one or more acoustic feature encodings using an acoustic feature decoder of the voice convertor.

A method obtains a first sound signal representative of a first sound, including a first spectrum envelope contour and a first reference spectrum envelope contour; obtains a second sound signal, representative of a second sound differing in sound characteristics from the first sound, including a second spectrum envelope contour and a second reference spectrum envelope contour; generates a synthesis spectrum envelope contour by transforming the first spectrum envelope contour based on a first difference between the first spectrum envelope contour and the first reference spectrum envelope contour at a first time point of the first sound signal, and a second difference between the second spectrum envelope contour and the second reference spectrum envelope contour at a second time point of the second sound signal; and generates a third sound signal representative of the first sound that has been transformed using the generated synthesis spectrum envelope contour.

A conversion learning device includes: a source encoding unit that converts, by using a first machine learning model, a feature amount sequence of a source domain that is a characteristic of conversion-source content data, into a first internal representation vector sequence that is a matrix in which internal representation vectors at individual locations of the feature amount sequence of the source domain are arranged; a target encoding unit that converts, by using a second machine learning model, a feature amount sequence of a target domain that is a characteristic of conversion-target content data, into a second internal representation vector sequence that is a matrix in which internal representation vectors at individual locations of the feature amount sequence of the target domain are arranged; an attention matrix calculation unit that calculates, by using the first internal representation vector sequence and the second internal representation vector sequence, an attention matrix that is a matrix mapping the individual locations of the feature amount sequence of the source domain to the individual locations of the feature amount sequence of the target domain, and calculates a third internal representation vector sequence that is a product of an internal representation vector sequence calculated by linear conversion of the first internal representation vector sequence and the attention matrix; a target decoding unit that calculates, by using the third internal representation vector sequence, a feature amount sequence of a conversion domain that is used to convert the source domain into the conversion domain, by using a third machine learning model; and a learning execution unit that causes at least one of the target encoding unit and the target decoding unit to learn such that a distance between a submatrix of the feature amount sequence of the target domain and a submatrix of the feature amount sequence of the conversion domain becomes shorter.

A conversion learning device includes: a source encoding unit that converts, by using a first machine learning model, a feature amount sequence of a source domain that is a characteristic of conversion-source content data, into a first internal representation vector sequence that is a matrix in which internal representation vectors at individual locations of the feature amount sequence of the source domain are arranged; a target encoding unit that converts, by using a second machine learning model, a feature amount sequence of a target domain that is a characteristic of conversion-target content data, into a second internal representation vector sequence that is a matrix in which internal representation vectors at individual locations of the feature amount sequence of the target domain are arranged; an attention matrix calculation unit that calculates, by using the first internal representation vector sequence and the second internal representation vector sequence, an attention matrix that is a matrix mapping the individual locations of the feature amount sequence of the source domain to the individual locations of the feature amount sequence of the target domain, and calculates a third internal representation vector sequence that is a product of an internal representation vector sequence calculated by linear conversion of the first internal representation vector sequence and the attention matrix; a target decoding unit that calculates, by using the third internal representation vector sequence, a feature amount sequence of a conversion domain that is used to convert the source domain into the conversion domain, by using a third machine learning model; and a learning execution unit that causes at least one of the target encoding unit and the target decoding unit to learn such that a distance between a submatrix of the feature amount sequence of the target domain and a submatrix of the feature amount sequence of the conversion domain becomes shorter.

An audio encoder for encoding an audio signal having a lower frequency band and an upper frequency band includes: a detector for detecting a peak spectral region in the upper frequency band of the audio signal; a shaper for shaping the lower frequency band using shaping information for the lower band and for shaping the upper frequency band using at least a portion of the shaping information for the lower band, wherein the shaper is configured to additionally attenuate spectral values in the detected peak spectral region in the upper frequency band; and a quantizer and coder stage for quantizing a shaped lower frequency band and a shaped upper frequency band and for entropy coding quantized spectral values from the shaped lower frequency band and the shaped upper frequency band.

An audio encoder for encoding an audio signal having a lower frequency band and an upper frequency band includes: a detector for detecting a peak spectral region in the upper frequency band of the audio signal; a shaper for shaping the lower frequency band using shaping information for the lower band and for shaping the upper frequency band using at least a portion of the shaping information for the lower band, wherein the shaper is configured to additionally attenuate spectral values in the detected peak spectral region in the upper frequency band; and a quantizer and coder stage for quantizing a shaped lower frequency band and a shaped upper frequency band and for entropy coding quantized spectral values from the shaped lower frequency band and the shaped upper frequency band.

A computer implemented method, system and computer program product are provided that implement a neural network (NN) audio filter. The method, system and computer program product obtain an electronic audio signal comprising a speech impaired message and apply the audio signal to the NN audio filter to modify the speech impaired message to form an unimpaired message. The method, system and computer program product output the unimpaired message.

A device for speech generation includes one or more processors configured to receive one or more control parameters indicating target speech characteristics. The one or more processors are also configured to process, using a multi-encoder, an input representation of speech based on the one or more control parameters to generate encoded data corresponding to an audio signal that represents a version of the speech based on the target speech characteristics.