Raw speech speaker-recognition

An apparatus comprising:

• a multi time-frequency resolution convolution neural network module having an input and an output;
• a two dimensional convolution neural network layers module having an input and an output;
• a discriminative fully-connected classifier layers module having an input and an output;
• wherein the multi time-frequency resolution convolution neural network module is configured to receive a raw speech signal from a human speaker at the input of the multi time-frequency resolution convolution neural network module;
• wherein the multi time-frequency resolution convolution neural network module is configured to process the raw speech signal to provide a first processed output at the output of the multi time-frequency resolution convolution neural network module;
• wherein the two dimensional convolution neural network layers module is configured to receive the first processed output at the input of the two dimensional convolution neural network layers module;
• wherein the two dimensional convolution neural network layers module is configured to process the first processed output to provide a second processed output at the output of the two dimensional convolution neural network layers module;
• wherein the discriminative fully-connected classifier layers module is configured to receive the second processed output at the input of the discriminative fully-connected classifier layers module;
• wherein the discriminative fully-connected classifier layers module is configured to process the second processed output to provide a third processed output at the output of the discriminative fully-connected classifier layers module;
• wherein one or more computer processors and one or more computer memories are provided for a combination of the multi time-frequency resolution convolution neural network module, the two dimensional convolution neural network layers module, and the discriminative fully-connected classifier layers module, to process the raw speech signal into the first processed output, to process the first processed output into the second processed output, and to process the second processed output into the third processed output, in accordance with computer software stored in the one or more computer memories as executed by the one or more computer processors;
• wherein the multi time-frequency resolution convolution neural network module processes the raw speech signal to provide a first processed output by performing a multi time-frequency analysis on a waveform of the raw speech signal using variable-sized kernels in a first convolution layer and thereby creating one or more feature maps that correspond to one or more spectrographs, each of the one or more spectrographs equivalent to a filter-bank analysis with variable kernel convolving filter sizes, wherein the one or more feature maps are the first processed output;
• wherein the two dimensional convolution neural network layers module processes the first processed output to provide the second processed output by processing the first processed output of one or more feature maps generated by the first convolution layer of the multi time-frequency resolution convolution neural network as a stack of one or more spectrographs, and subjecting the stack of one or more spectrographs to max-pooling to reduce the stack of one or more spectrographs to a feature-map of constant size and further processed by one or more convolutional layers of the two dimensional convolution neural network layers module, viewing the stack of one or more spectrographs, as a volume of time-frequency data and applying a pipeline of four image convolutional neural network layers of the two dimensional convolution neural network on the stack of one or more spectrographs as a flattened image convolution, which results in a low dimensional embedding vector, which is flattened and which is provided as the second processed output;
• wherein the discriminative fully-connected classifier layers module processes the second processed output to provide a third processed output by performing a multi-way speaker-identification task by yielding multiple posterior probabilities over multiple speaker classes;
• and
• wherein the discriminative fully-connected classifier layers module processes the second processed output to provide a third processed output by performing a two way speaker-verification task by respectively yielding one or more posterior probabilities over one or more target and non target speaker classes.

A method comprising:

• receiving a raw speech signal from a human speaker;
• processing the raw speech signal by use of a computer processor in accordance with computer software stored in computer memory to provide a first processed output;
• processing the first processed output by use of a computer processor in accordance with computer software stored in computer memory to provide a second processed output;
• processing the second processed output by use of a computer processor in accordance with computer software stored in computer memory to provide a third processed output;
• wherein the raw speech signal is processed to provide the first processed output by performing a multi time-frequency analysis on a waveform of the raw speech signal using variable-sized kernels in a first convolution layer and thereby creating one or more feature maps that correspond to one or more spectrographs, each of the one or more spectrographs equivalent to a filter-bank analysis with variable kernel convolving filter sizes, wherein the one or more feature maps are the first processed output;
• wherein the first processed output is processed to provide the second processed output by processing the one or more feature maps of the first processed output as a stack of one or more spectrographs, and subjecting the stack of one or more spectrographs to max-pooling to reduce the stack of one or more spectrographs to a feature-map of constant size and then the feature-map of constant size is further processed by one or more convolutional layers, viewing the stack of one or more spectrographs, as a volume of time-frequency data and applying a pipeline of four image convolutional neural network layers on the stack of one or more spectrographs as a flattened image convolution, which results in a low dimensional embedding vector, which is flattened and which is provided as the second processed output;
• and
• wherein the second processed output is processed to provide the third processed output by performing a multi-class speaker-identification task on the second processed output by respectively yielding one or more posterior probabilities over one or more speaker classes.

A method comprising

• receiving a raw speech signal from a human speaker;
• processing the raw speech signal using a computer processor in accordance with computer software stored in computer memory to provide a first processed output;
• processing the first processed output using a computer processor in accordance with computer software stored in computer memory to provide a second processed output;
• processing the second processed output in accordance with computer software stored in computer memory to provide a third processed output;
• wherein the raw speech signal is processed to provide the first processed output by performing a multi time-frequency analysis on a waveform of the raw speech signal using variable-sized kernels in a first convolution layer and thereby creating one or more feature maps that correspond to one or more spectrographs, each of the one or more spectrographs equivalent to a filter-bank analysis with variable kernel convolving filter sizes, wherein the one or more feature maps are the first processed output;
• wherein the first processed output is processed to provide the second processed output by processing the one or more feature maps of the first processed output as a stack of one or more spectrographs, and subjecting the stack of one or more spectrographs to max-pooling to reduce the stack of one or more spectrographs to a feature-map of constant size and then the feature-map of constant size is further processed by one or more convolutional layers, viewing the stack of one or more spectrographs, as a volume of time-frequency data and applying a pipeline of four image convolutional neural network layers on the stack of one or more spectrographs as a flattened image convolution, which results in a low dimensional embedding vector, which is flattened and which is provided as the second processed output;
• and
• wherein the second processed output is processed to provide the third processed output by performing a two way speaker-verification task by respectively yielding one or more posterior probabilities over one or more target and non-target speaker classes.