A system and method are presented for the correction of packet loss in audio in automatic speech recognition (ASR) systems. Packet loss correction, as presented herein, occurs at the recognition stage without modifying any of the acoustic models generated during training. The behavior of the ASR engine in the absence of packet loss is thus not altered. To accomplish this, the actual input signal may be rectified, the recognition scores may be normalized to account for signal errors, and a best-estimate method using information from previous frames and acoustic models may be used to replace the noisy signal.

The present disclosure is generally directed a system to detect activation phrases within input audio signals transmitted over a low-bandwidth network. The system can use a two-stage activation phrase detection process. First a sensing device, which can include a plurality of microphones for detecting an input audio signal, can detect an input audio signal that includes a candidate activation phrase. Second, the sensing device can transmit the recordings of the input audio signal to a client device for confirmation that the input audio signal includes the activation phrase.

A sound signal processing system includes: a sound signal processing apparatus executing non-linear signal processing on a collected sound signal collected by a microphone, and transmitting, to an information processing apparatus, both a pre-execution sound signal before the non-linear signal processing is executed and a post-execution sound signal after the non-linear signal processing is executed; and the information processing apparatus receiving the pre-execution sound signal and the post-execution sound signal from the sound signal processing apparatus, and executing first processing on the pre-execution sound signal and executing second processing on the post-execution sound signal, the second processing being different from the first processing.

A sound signal processing system includes: a sound signal processing apparatus executing non-linear signal processing on a collected sound signal collected by a microphone, and transmitting, to an information processing apparatus, both a pre-execution sound signal before the non-linear signal processing is executed and a post-execution sound signal after the non-linear signal processing is executed; and the information processing apparatus receiving the pre-execution sound signal and the post-execution sound signal from the sound signal processing apparatus, and executing first processing on the pre-execution sound signal and executing second processing on the post-execution sound signal, the second processing being different from the first processing.

A method and apparatus for generating a speaker identification neural network include generating a first neural network that is trained to identify a first speaker with respect to a first voice signal in a first environment, generating a second neural network for identifying a second speaker with respect to a second voice signal in a second environment, and generating the speaker identification neural network by training the second neural network based on a teacher-student training model in which the first neural network is set to a teacher neural network and the second neural network is set to a student neural network.

A method and apparatus for generating a speaker identification neural network include generating a first neural network that is trained to identify a first speaker with respect to a first voice signal in a first environment, generating a second neural network for identifying a second speaker with respect to a second voice signal in a second environment, and generating the speaker identification neural network by training the second neural network based on a teacher-student training model in which the first neural network is set to a teacher neural network and the second neural network is set to a student neural network.

A target speech signal extraction method for robust speech recognition includes: initializing a steering vector for a target speech source and an adaptive vector, setting a real output channel of the target speech source as an output by the adaptive vector, initializing adaptive vectors for a noise and setting a dummy channel as an output by the adaptive vectors for the noise; setting a cost function for minimizing dependency between a real output for the target speech source and a dummy output for the noise; setting an auxiliary function to the cost function, and updating the adaptive vector for the target speech source and the adaptive vectors for the noise by using the auxiliary function and the steering vector; estimating the target speech signal by using the adaptive vector thereby extracting the target speech signal from the input signals; and updating the steering vector for the target speech source.

A target speech signal extraction method for robust speech recognition includes: initializing a steering vector for a target speech source and an adaptive vector, setting a real output channel of the target speech source as an output by the adaptive vector, initializing adaptive vectors for a noise and setting a dummy channel as an output by the adaptive vectors for the noise; setting a cost function for minimizing dependency between a real output for the target speech source and a dummy output for the noise; setting an auxiliary function to the cost function, and updating the adaptive vector for the target speech source and the adaptive vectors for the noise by using the auxiliary function and the steering vector; estimating the target speech signal by using the adaptive vector thereby extracting the target speech signal from the input signals; and updating the steering vector for the target speech source.

A system accesses a first digital audio file that includes a plurality of spoken instructions. The system converts the first digital audio file to a first spectrogram image, applies a filter to determine whether an image quality of the first spectrogram image is below a predetermined image quality, and in response, generates a second spectrogram image from the first spectrogram image using a training model. The system converts the second spectrogram image to a second digital audio file and converts the second digital audio file into multiple vectors that each correspond to a particular spoken instruction. The system identifies related vectors and concatenates the related vectors together in order to create a plurality of concatenated vectors. The system generates, using the plurality of concatenated vectors, a third digital audio file that includes concatenated spoken instructions from the first digital audio file.

A system accesses a first digital audio file that includes a plurality of spoken instructions. The system converts the first digital audio file to a first spectrogram image, applies a filter to determine whether an image quality of the first spectrogram image is below a predetermined image quality, and in response, generates a second spectrogram image from the first spectrogram image using a training model. The system converts the second spectrogram image to a second digital audio file and converts the second digital audio file into multiple vectors that each correspond to a particular spoken instruction. The system identifies related vectors and concatenates the related vectors together in order to create a plurality of concatenated vectors. The system generates, using the plurality of concatenated vectors, a third digital audio file that includes concatenated spoken instructions from the first digital audio file.