Evaluating reliability of audio data for use in speech processing

A method comprising:

• receiving, by processing circuitry having access to a storage device configured to store a machine learning model trained with audio feature values to determine a reliability of an audio segment for performing speech processing, an audio dataset comprising a sequence of audio segments including speech from an unknown speaker;
• extracting, by the processing circuitry for each audio segment of the sequence of audio segments, a set of audio feature values corresponding to a set of audio features, wherein each audio feature value of the set of audio feature values indicates a prevalence of the corresponding audio feature of the set of audio features in the audio segment;
• executing, by the processing circuitry, the machine learning model to determine, for each audio segment of the sequence of audio segments, a reliability score based on the set of audio feature values corresponding to the respective audio segment, wherein the reliability score indicates a reliability of the audio segment for performing speech processing;
• selecting, by the processing circuitry, based on the reliability score for each audio segment, one or more audio segments from the sequence of audio segments;
• storing, by the processing circuitry, the one or more audio segments as most valuable for performing speaker identification;
• determining, by the processing circuitry, based on the one or more audio segments, a probability indicating whether the speech from the unknown speaker corresponds to a target speaker;
• and
• outputting, by the processing circuitry, an indication of the probability indicating whether the speech from the unknown speaker corresponds to the target speaker.

Non-transitory computer-readable media medium comprising instructions that, when executed by a processor, cause the processor to:

• receive an audio dataset comprising a sequence of audio segments including speech from an unknown speaker;
• extract, for each audio segment of the sequence of audio segments, a set of audio feature values corresponding to a set of audio features, wherein each audio feature value of the set of audio feature values indicates a prevalence of the corresponding audio feature of the set of audio features in the audio segment;
• execute a machine learning model to determine, for each audio segment of the sequence of audio segments, a reliability score based on the set of audio feature values corresponding to the respective audio segment, wherein the reliability score indicates a reliability of the audio segment for performing speech processing;
• select, based on the reliability score for each audio segment, one or more audio segments from the sequence of audio segments;
• store the one or more audio segments as most valuable for performing speaker identification;
• determine, based on the one or more audio segments, a probability indicating whether the speech from the unknown speaker corresponds to a target speaker;
• and
• output an indication of the probability indicating whether the speech from the unknown speaker corresponds to the target speaker.

A computing system comprising:

• a storage device configured to store a machine learning model and training data comprising a plurality of training audio datasets, wherein each training audio dataset of the plurality of training audio datasets is known to include speech from a particular target class, wherein each training audio dataset of the plurality of training audio datasets comprises a sequence of training audio segments;
• and
• processing circuitry having access to the storage device, configured to determine a probability indicating whether speech from an unknown speaker corresponds to a target speaker, and configured to train the machine learning model to determine a reliability of an audio segment for performing speech processing,
• wherein to train the machine learning model, the processing circuitry is configured to:
• execute a speech processing model to perform a plurality of speech processing determinations, wherein each speech processing determination of the plurality of speech processing determinations involves determining whether a training audio segment of a training audio dataset corresponds to a particular target class;
• determine, based on the particular target class known to be associated with each training audio dataset of the plurality of training audio datasets, a quality of each speech processing determination of the plurality of speech processing determinations;
• identify one or more audio features of a plurality of audio features present in the training audio segment corresponding to each speech processing determination of the plurality of speech processing determinations;
• and
• select a set of audio features of the plurality of audio features based on the quality of each speech processing determination of the plurality of speech processing determinations and the one or more audio features of the plurality of audio features present in the audio segment corresponding to each speech processing determination of the plurality of speech processing determinations, and
• wherein to determine the probability indicating whether speech from the unknown speaker corresponds to the target speaker, the processing circuitry is configured to:
• receive an audio dataset comprising a sequence of audio segments including speech from the unknown speaker;
• extract, for each audio segment of the sequence of audio segments, a set of audio feature values corresponding to the set of audio features, wherein each audio feature value of the set of audio feature values indicates a prevalence of the corresponding audio feature of the set of audio features in the audio segment;
• execute the machine learning model to determine, for each audio segment of the sequence of audio segments, a reliability score based on the set of audio feature values corresponding to the respective audio segment, wherein the reliability score indicates a reliability of the audio segment for performing speech processing;
• select, based on the reliability score for each audio segment, one or more audio segments from the sequence of audio segments;
• store the one or more audio segments as most valuable for performing speaker identification;
• determine, based on the one or more audio segments, a probability indicating whether the speech from the unknown speaker corresponds to the target speaker;
• and
• output an indication of the probability indicating whether the speech from the unknown speaker corresponds to the target speaker.