Voice/non-voice determination device, voice/non-voice determination model parameter learning device, voice/non-voice determination method, voice/non-voice determination model parameter learning method, and program

A voice/non-voice determination model parameter learning device comprising:

• processing circuitry configured to:
• receive a set of pair data of an acoustic signal and acoustic scene label which is a label of a scene where the acoustic signal is collected as first training data;
• receive a set of pair data of the acoustic signal and the acoustic signal after speech enhancement as second training data;
• receive a set of pair data of the acoustic signal and a label indicating a voice/non-voice state as third training data;
• learn a parameter of a first model which is a model for acoustic scene classification using the first training data;
• learn a parameter of a second model which is a model for speech enhancement using the second training data;
• learn a parameter of a third model which is a model for voice/non-voice determination using the third training data, for fourth training data including a total of four pieces of information of acoustic scene information which is output of the first model with respect to the acoustic signal included in the third training data, speech enhancement information which is output of the second model with respect to the acoustic signal included in the third training data, the label included in the third training data, and the acoustic signal included in the third training data:
• learn a parameter of a fourth model which is a model for phoneme recognition using a set of pair data of the acoustic signal and a phoneme label of the acoustic signal as fifth training data;
• learn a parameter of a fifth model which is a model for speaker recognition using a set of pair data of the acoustic signal and a speaker label of the acoustic signal as sixth training data,
• wherein the processing circuitry learns the parameter of the third model for seventh training data including a total of six pieces of information of the acoustic scene information, the speech enhancement information, the label included in the third training data, the acoustic signal included in the third training data, phoneme recognition information which is output of the fourth model with respect to the acoustic signal included in the third training data, and speaker recognition information which is output of the fifth model with respect to the acoustic signal included in the third training data;
• and
• cause outputting a result of speech recognition data associated with the acoustic signal according to the voice/non-voice label as output from the third model and the phoneme recognition information as output from the fourth model.