Self-learning neuromorphic acoustic model for speech recognition
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for recognizing speech using a spiking neural network acoustic model implemented on a neuromorphic processor are described. In one aspect, a method includes receiving, a trained acoustic model implemented as a spiking neural network (SNN) on a neuromorphic processor of a client device, a set of feature coefficients that represent acoustic energy of input audio received from a microphone communicably coupled to the client device. The acoustic model is trained to predict speech sounds based on input feature coefficients. The acoustic model generates output data indicating predicted speech sounds corresponding to the set of feature coefficients that represent the input audio received from the microphone. The neuromorphic processor updates one or more parameters of the acoustic model using one or more learning rules and the predicted speech sounds of the output data.
1. A method for recognizing speech, the method comprising:
receiving, a trained acoustic model implemented as a spiking neural network (SNN) on a neuromorphic processor of a client device, a set of feature coefficients that represent acoustic energy of input audio received from a microphone communicably coupled to the client device, wherein the acoustic model is trained to predict speech sounds based on input feature coefficients;
generating, by the acoustic model, output data indicating predicted speech sounds corresponding to the set of feature coefficients that represent the input audio received from the microphone;
updating, by the neuromorphic processor, one or more parameters of the acoustic model using one or more learning rules and the predicted speech sounds of the output data, wherein at least one learning rule is configured to update parameters of the acoustic model based on each speech recognition event for which the acoustic model generates a prediction of speech sounds; and
initiating an action based on the output data.
2. The method of claim 1 , wherein initiating the action comprises sending the output data to a cloud-based platform comprising a lexicon that is trained to output predicted words or phrases based on the predicted speech sounds, which are input into a language model trained to output predicted sentences or phrases, wherein the cloud-based platform determines the action based on the predicted sentences or phrases and causes the action to be performed.
3. The method of claim 1 , wherein at least one of the one or more learning rules is configured to update the parameters of the acoustic model continuously and/or in real-time.
4. The method of claim 1 , wherein the one or more learning rules comprise a user correction learning rule configured to update the parameters of the acoustic model based on an error between the predicted speech sounds and corrected speech sounds corresponding to a user correction of words or phrases output based on the predicted speech sounds.
5. The method of claim 1 , wherein the one or more learning rules comprise a sound context learning rule configured to reduce a difference between predicted speech sounds output the acoustic model and an inferred class of the predicted speech sounds based on context.
6. The method of claim 1 , wherein the one or more learning rules comprise a sound prediction learning rule configured to reduce an error between a prediction of upcoming speech sounds to be generated by the acoustic model and corresponding future predicted speech sounds output by the acoustic model.
7. The method of claim 1 , wherein the one or more learning rules comprise a sound distribution learning rule configured to reduce a difference between an observed distribution of speech sounds predicted by the acoustic model and a distribution of speech sounds in a reference dataset.
8. The method of claim 1 , further comprising receiving an initial trained version of the acoustic model from a cloud-based platform and configuring the SNN based on the initial trained version of the acoustic model.
9. The method of claim 1 , wherein initiating the action comprises providing the output data to a lexicon that is trained to output predicted words or phrases based on the predicted speech sounds, which are input into a language model trained to output predicted sentences or phrases wherein the client device determines the action based on the predicted sentences or phrases and causes the action to be performed.
10. A system comprising:
one or more processors including a neuromorphic processor; and
one or more storage devices storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
receiving, a trained acoustic model implemented as a spiking neural network (SNN) on the neuromorphic processor, a set of feature coefficients that represent acoustic energy of input audio received from the microphone, wherein the acoustic model is trained to predict speech sounds based on input feature coefficients;
generating, by the acoustic model, output data indicating predicted speech sounds corresponding to the set of feature coefficients that represent the input audio received from the microphone;
updating one or more parameters of the acoustic model using one or more learning rules and the predicted speech sounds of the output data, wherein at least one learning rule is configured to update parameters of the acoustic model based on each speech recognition event for which the acoustic model generates a prediction of speech sounds; and
initiating an action based on the output data.
11. The system of claim 10 , wherein initiating the action comprises sending the output data to a cloud-based platform comprising a lexicon that is trained to output predicted words or phrases based on the predicted speech sounds, which are input into a language model trained to output predicted sentences or phrases, wherein the cloud-based platform determines the action based on the predicted sentences or phrases and causes the action to be performed.
12. The system of claim 10 , wherein at least one of the one or more learning rules is configured to update the parameters of the acoustic model continuously and/or in real-time.
13. The system of claim 10 , wherein the one or more learning rules comprise a user correction learning rule configured to update the parameters of the acoustic model based on an error between the predicted speech sounds and corrected speech sounds corresponding to a user correction of words or phrases output based on the predicted speech sounds.
14. The system of claim 10 , wherein the one or more learning rules comprise a sound context learning rule configured to reduce a difference between predicted speech sounds output the acoustic model and an inferred class of the predicted speech sounds based on context.
15. The system of claim 10 , wherein the one or more learning rules comprise a sound prediction learning rule configured to reduce an error between a prediction of upcoming speech sounds to be generated by the acoustic model and corresponding future predicted speech sounds output by the acoustic model.
16. The system of claim 10 , wherein the one or more learning rules comprise a sound distribution learning rule configured to reduce a difference between an observed distribution of speech sounds predicted by the acoustic model and a distribution of speech sounds in a reference dataset.
17. The system of claim 10 , wherein the operations comprise receiving an initial trained version of the acoustic model from a cloud-based platform and configuring the SNN based on the initial trained version of the acoustic model.
18. The system of claim 10 , wherein initiating the action comprises providing the output data to a lexicon that is trained to output predicted words or phrases based on the predicted speech sounds, which are input into a language model trained to output predicted sentences or phrases wherein the client device determines the action based on the predicted sentences or phrases and causes the action to be performed.
19. A non-transitory computer readable medium storing instructions that, when executed by one or more processors including a neuromorphic processor, cause the one or more processors to perform operations comprising:
receiving, a trained acoustic model implemented as a spiking neural network (SNN) on the neuromorphic processor, a set of feature coefficients that represent acoustic energy of input audio received from the microphone, wherein the acoustic model is trained to predict speech sounds based on input feature coefficients;
generating, by the acoustic model, output data indicating predicted speech sounds corresponding to the set of feature coefficients that represent the input audio received from the microphone;
updating one or more parameters of the acoustic model using one or more learning rules and the predicted speech sounds of the output data, wherein at least one learning rule is configured to update parameters of the acoustic model based on each speech recognition event for which the acoustic model generates a prediction of speech sounds; and
initiating an action based on the output data.
20. The non-transitory computer readable medium of claim 19 , wherein initiating the action comprises sending the output data to a cloud-based platform comprising a lexicon that is trained to output predicted words or phrases based on the predicted speech sounds, which are input into a language model trained to output predicted sentences or phrases, wherein the cloud-based platform determines the action based on the predicted sentences or phrases and causes the action to be performed.