Augmenting voice samples based on distributions of speaker classes

A non-transitory computer-readable storage medium storing instructions that when executed by a processor cause the processor to perform steps comprising:

• obtaining a plurality of voice samples;
• grouping the plurality of voice samples into at least one class of spectral representations amongst a plurality of classes of spectral representations comprising a first class of spectral representations and a second class of spectral representations;
• obtaining warp distributions associated with the plurality of classes of spectral representations;
• determining one or more spectral change ratios based on a comparison of the warp distributions comprising one or more of,
• determining a first spectral change ratio for centering a warp distribution of the first class around a peak warp value associated with the second class of spectral representations;
• and
• determining a second spectral change ratio for centering a warp distribution of the second class around a peak warp value associated with the first class of spectral representations;
• determining one or more transformations based at least in part on the one or more spectral change ratios;
• applying the one or more transformations to the plurality of voice samples grouped into the at least one class of spectral representations to generate a set of augmented voice samples;
• compiling a training dataset using at least the set of augmented voice samples;
• training a machine learning model using the training dataset compiled using at least the set of augmented voice samples, wherein the machine learning model is an acoustic model;
• and
• receiving speech data, applying the machine learning model to map a relationship between the speech data and a plurality of linguistic units, and generating a transcription of the speech data based on the plurality of linguistic units.

A system comprising:

• a processor;
• and
• a non-transitory computer-readable storage medium storing instructions that when executed by the processor cause the processor to perform steps comprising:
• receiving, from a client device, a request to recognize speech data;
• applying a machine learning model to map a relationship between the speech data and a plurality of linguistic units to generate a transcription of the speech data based on the plurality of linguistic units, wherein the machine learning model is an acoustic model trained using a training dataset compiled from at least a set of augmented voice samples;
• and
• sending a response to the client device, wherein the set of augmented voice samples are generated by:
• grouping a plurality of voice samples into a plurality of classes of spectral representations comprising a first class of spectral representations and a second class of spectral representations;
• determining spectral change ratios based on a comparison of warp distributions associated with the plurality of classes of spectral representations comprising one or more of,
• determining a first spectral change ratio for centering a warp distribution of the first class around a peak warp value associated with the second class of spectral representations;
• and
• determining a second spectral change ratio for centering a warp distribution of the second class around a peak warp value associated with the first class of spectral representations;
• applying transformations to the plurality of voice samples based at least on the spectral change ratios to generate the set of augmented voice samples;
• and
• compiling the training dataset using at least the set of augmented voice samples.