Low resource key phrase detection for wake on voice

A system for performing key phrase detection comprising:

• a memory configured to store an acoustic model, a start state based rejection model, and a key phrase model associated with a predetermined key phrase;
• and
• a digital signal processor coupled to the memory, the digital signal processor to generate, based on the acoustic model, a time series of scores of sub-phonetic units based on a time series of feature vectors representative of an audio input, to update the start state based rejection model and the key phrase model based on at least some of the time series of scores of sub-phonetic units to generate a rejection likelihood score and a key phrase likelihood score, wherein the start state based rejection model consists of a single rejection state and comprises a plurality of rejection model self loops each associated with a particular score of the scores of sub-phonetic units of the acoustic model, wherein the key phrase model comprises a plurality of key phrase states interconnected by transitions therebetween, the plurality of key phrase states each comprising a self loop associated with a particular score of the scores of sub-phonetic units of the acoustic model, and wherein the start state based rejection model and the key phrase model are connected by a first transition from the single rejection state to a first key phrase state of the plurality of key phrase states, and to determine whether the received audio input is associated with the predetermined key phrase based on the rejection likelihood score and the key phrase likelihood score.

At least one non-transitory machine readable medium comprising a plurality of instructions that, in response to being executed on a device, cause the device to generate a key phrase detection model comprising a start state based rejection model, a key phrase model, and a pruned acoustic model by:

• training an acoustic model having a plurality of output nodes, the output nodes comprising a plurality of sub-phonetic units in the form of tied context-dependent triphone HMM-states, wherein each of the tied context-dependent triphone HMM-states is associated with one of a plurality of monophones;
• and
• generating a selected subset of the output nodes by:
• determining a usage rate for each of the sub-phonetic units during the training;
• including, in the selected subset, at least one output node corresponding to a highest usage rate sub-phonetic unit for each of the plurality of monophones;
• including, in the selected subset, output nodes corresponding to states of the key phrase model;
• and
• discarding remaining output nodes not in the selected subset;
• and
• generating a pruned acoustic model having outputs comprising the selected subset of the output nodes, wherein the start state based rejection model comprises rejection model self loops associated with first outputs of the pruned acoustic model, the states of the key phrase model are associated with second outputs of the acoustic model, and wherein the start state based rejection model and the key phrase model are connected by a transition from a single rejection state of the start state based rejection model to a first state of the states of the key phrase model.