Device for predicting voice conversion model, method of predicting voice conversion model, and computer program product
According to an embodiment, a voice processing device includes an interface system, a determining processor, and a predicting processor. The interface system configured to receive neutral voice data representing audio in a neutral voice of a user. The determining processor configured to determine a predictive parameter based at least in part on the neutral voice data. The predicting processor configured to predict a voice conversion model for converting the neutral voice of the speaker to a target voice using at least the predictive parameter.
1. A device for predicting a voice conversion model, the device comprising:
an interface system configured to receive neutral voice data representing audio in a neutral voice of a user;
a determining processor, implemented in computer hardware, configured to determine a predictive parameter based at least in part on the neutral voice data; and
a predicting processor, implemented in computer hardware, configured to predict a voice conversion model for converting the neutral voice of the speaker to a target voice tone using at least the predictive parameter, wherein
a plurality of neutral voice predictive models are respectively associated with voice conversion predictive models each of which is optimized for converting the corresponding neutral voice predictive model to a voice model of the target voice,
the neutral voice data comprises acoustic feature quantity data representing a feature of the voice obtained by analyzing the audio in the neutral voice of the user and language attribute date representing an attribute of a language obtained by analyzing the audio in the neutral voice of the user, and
the determining processor is configured to:
calculate a likelihood of a linear sum of a vector based at least in part on the neutral voice predictive models with respect to the acoustic feature quantity data and the language attribute data,
determine, as a weight, a coefficient of the linear sum comprising the highest calculated likelihood, and
determine the predictive parameter generated by adding, to a model parameter of each voice conversion predictive model, the weight determined with respect to the corresponding neutral voice predictive model.
2. A method of predicting a voice conversion model, the method comprising:
receiving, by an interface system, neutral voice data representing audio in a calm voice tone of a user;
determining, by a determining processor implemented in computer hardware, a predictive parameter based at least in part on the neutral voice data; and
predicting, by a predicting processor implemented in computer hardware, a voice conversion model for converting the neutral voice of the speaker to a target voice using at least the predictive parameter, wherein
a plurality of neutral voice predictive models are respectively associated with voice conversion predictive models each of which is optimized for converting the corresponding neutral voice predictive model to a voice model of the target voice,
the neutral voice data comprises acoustic feature quantity data representing a feature of the voice obtained by analyzing the audio in the neutral voice of the user and language attribute date representing an attribute of a language obtained by analyzing the audio in the neutral voice of the user, and
the determining includes:
calculating a likelihood of a linear sum of a vector based at least in part on the neutral voice predictive models with respect to the acoustic feature quantity data and the language attribute data,
determining, as a weight, a coefficient of the linear sum comprising the highest calculated likelihood, and
determining the predictive parameter generated by adding, to a model parameter of each voice conversion predictive model, the weight determined with respect to the corresponding neutral voice predictive model.
3. A computer program product comprising a non-transitory computer-readable medium containing a computer program that causes a computer to function as:
an interface system configured to receive neutral voice data representing audio in a neutral voice of a user;
a determining processor configured to determine a predictive parameter at least in part on the neutral voice data; and
a predicting processor configured to predict a voice conversion model for converting the neutral voice of the speaker to a target voice, wherein
a plurality of neutral voice predictive models are respectively associated with voice conversion predictive models each of which is optimized for converting the corresponding neutral voice predictive model to a voice model of the target voice,
the neutral voice data comprises acoustic feature quantity data representing a feature of the voice obtained by analyzing the audio in the neutral voice of the user and language attribute date representing an attribute of a language obtained by analyzing the audio in the neutral voice of the user, and
the determining processor is configured to:
calculate a likelihood of a linear sum of a vector based at least in part on the neutral voice predictive models with respect to the acoustic feature quantity data and the language attribute data,
determine, as a weight, a coefficient of the linear sum comprising the highest calculated likelihood, and
determine the predictive parameter generated by adding, to a model parameter of each voice conversion predictive model, the weight determined with respect to the corresponding neutral voice predictive model.