Systems, methods, and computer-readable storage media for audio signal processing

A computer-readable storage medium storing computer-readable instructions executable by one or more processors, that when executed by the one or more processors, cause the one or more processors to perform operations comprising:

• receiving an audio signal;
• dividing the audio signal into a plurality of frames using a predetermined window size and a predetermined window shift;
• and
• grouping the plurality of frames into a plurality of segments, a respective segment of the plurality of segments including a predetermined number of frames;
• obtaining a plurality of multi-dimensional features based on the audio signal;
• wherein obtaining the plurality of multi-dimensional features based on the audio signal includes converting the plurality of frames into a plurality of multi-dimensional filter bank features;
• normalizing the plurality of multi-dimensional filter bank features;
• obtaining a plurality of segment-level representations based on the plurality of multi-dimensional features;
• obtaining, by an utterance-level embedding extraction mechanism, an utterance-level representation based on the plurality of segment-level representations, wherein the utterance-level embedding extraction mechanism is unfolded according to a series of time steps, and one or more jump connections are added between two time steps of the series of time steps in a time direction of the utterance-level embedding extraction mechanism;
• and
• recognizing a speaker from the audio signal based on the utterance-level representation.

A system comprising:

• one or more processors;
• and
• memory communicatively coupled to the one or more processors, the memory storing computer-executable modules executable by the one or more processors, the computer-executable modules including:
• a data preparation module configured to receive an audio signal and obtain a plurality of multi-dimensional features based on the audio signal, wherein the data preparation module is further configured to:
• divide the audio signal into a plurality of frames using a predetermined window size and a predetermined window shift;
• group the plurality of frames into a plurality of segments, a respective segment of the plurality of segments including a predetermined number of frames;
• convert the plurality of frames into a plurality of multi-dimensional filter bank features;
• and
• normalizing the plurality of multi-dimensional filter bank features;
• a segment-level embedding extraction module configured to obtain a plurality of segment-level representations based on the plurality of multi-dimensional features;
• an utterance-level embedding extraction module configured to obtain an utterance-level representation based on the plurality of segment-level representations, wherein the utterance-level embedding extraction module is configured to be unfolded according to a series of time steps, and one or more jump connections are added between two time steps of the series of time steps in a time direction of the utterance-level embedding extraction module;
• and
• a classification module configured to recognize a speaker from the audio signal based on the utterance-level representation by classifying the utterance-level representation into a class associated with an ID of the speaker.