Voice conversion using deep neural network with intermediate voice training

A method of converting source speech to target speech, the method comprising:

• receiving a first plurality of target voice training data, wherein the first plurality of target voice training data is speech;
• extracting a first set of spectral features and pitch from the first plurality of target voice training data;
• retrieving, based on the first set of spectral features, a second set of spectral features for an intermediate speaker that match the first plurality of target voice training data;
• perform dynamic time warping to match the second set of spectral features and the first et of spectral features;
• training a neural network to convert the second set of spectral features to the first set of spectral features;
• acquiring with a mobile phone, a third plurality of source speech from a source speaker, wherein the source speech is speech;
• converting the third plurality of source speech to a fourth plurality of intermediate speech, wherein the fourth plurality of intermediate speech is speech;
• converting the fourth plurality of intermediate speech to a target speech sequence using the neural network, wherein the target speech sequence comprises a fifth plurality of target speech;
• convert the target speech sequence to target speech audio using the pitch from the first plurality of target voice training data, wherein the target speech sequence is speech;
• and
• playing, with the mobile phone, the target speech audio to the source speaker.

A method of converting source speech to target speech, the method comprising:

• receiving a first plurality of target voice training data, wherein the first plurality of target voice training data is speech;
• recognizing a plurality of words spoken by a target speaker from the first plurality of target voice training data;
• retrieving a second plurality of intermediate speech based on the recognized words;
• perform dynamic time warping to align the retrieve second plurality of intermediate speech with the corresponding first plurality of target voice training data;
• training a neural network to convert the aligned second plurality of intermediate speech to first plurality of target speech training data;
• acquiring, with a mobile phone, a third plurality of source speech from a source speaker;
• converting the third plurality of source speech to a fourth plurality of intermediate speech, wherein the fourth plurality of intermediate speech is speech;
• converting the fourth plurality of intermediate speech to a target speech sequence using the neural network, wherein the target speech sequence comprises a fifth plurality of target speech;
• converting the target speech sequence to target speech audio, wherein the target speech sequence is speech;
• and
• playing with the mobile phone, the target speech audio to the source speaker.