CPC G10L 15/063 (2013.01) | 20 Claims |
1. A computer-implemented method comprising:
training first parameters for a first speech processing model using a first dataset;
sending the first speech processing model to a remote device;
determining a second dataset, wherein the second dataset includes at least a portion of the first dataset;
determining a size constraint for a model update object, wherein the size constraint is determined such that processing the first speech processing model using the model update object generates a second speech processing model that achieves a desired accuracy;
using the size constraint, the first speech processing model, and the second dataset to train the model update object to include first values, wherein processing the first parameters using the first values generates second parameters for the second speech processing model;
sending the model update object to the remote device; and
causing the remote device to generate the second speech processing model using the first speech processing model and the model update object.
|