Training multiple neural networks with different accuracy
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a deep neural network. One of the methods includes generating a plurality of feature vectors that each model a different portion of an audio waveform, generating a first posterior probability vector for a first feature vector using a first neural network, determining whether one of the scores in the first posterior probability vector satisfies a first threshold value, generating a second posterior probability vector for each subsequent feature vector using a second neural network, wherein the second neural network is trained to identify the same key words and key phrases and includes more inner layer nodes than the first neural network, and determining whether one of the scores in the second posterior probability vector satisfies a second threshold value.
1. A computer-implemented method that when executed on data processing hardware of a user device causes the data processing hardware to perform operations comprising:
receiving a digital representation of speech recorded by a microphone of the user device;
generating, using a first neural network executing on the user device, a probability score representing a likelihood that a specific keyword or key phrase is present in the digital representation of speech, the first neural network trained to identify the specific keyword or key phrase;
determining whether the probability score satisfies a probability score threshold; and
in response to determining that the probability score satisfies the probability score threshold:
generating, using a second neural network executing on the user device, a classification result confirming whether or not the specific keyword or key phrase is present in the digital representation of speech, the second neural network trained to identify the specific keyword or key phrase and having a higher accuracy than the first neural network; and
in response to the classification result confirming that the specific keyword or key phrase is present in the digital representation of speech, activating the user device from a standby mode to perform a task.
2. The computer-implemented method of claim 1 , wherein the digital representation of speech comprises a stream of fixed-length audio frames.
3. The computer-implemented method of claim 1 , wherein the first neural network executing on the user device requires less computation than the second neural network executing on the user device.
4. The computer-implemented method of claim 1 , wherein:
the classification result comprises a posterior probability score representing the likelihood that the keyword or key phrase is present in the digital representation of speech; and
the classification result confirms that the specific keyword or key phrase is present in the digital representation of speech when the posterior probability score satisfies a posterior probability score threshold.
5. The computer-implemented method of claim 1 , wherein:
the first neural network is trained to identify the specific keyword or key phrase using a first training set for a first quantity of iterations; and
the second neural network is trained to identify the specific keyword or key phrase using a second training set for a second quantity of iterations, the second quantity of iterations greater than the first quantity of iterations.
6. The computer-implemented method of claim 5 , wherein the second training set and the first training set comprise the same training set.
7. The computer-implemented method of claim 1 , wherein:
the first neural network comprises a first quantity of nodes; and
the second neural network comprises a second quantity of nodes, the second quantity of nodes greater than the first quantity of nodes.
8. The computer-implemented method of claim 1 , wherein the first neural network comprises a higher false positive rate than the second neural network.
9. The computer-implemented method of claim 1 , wherein the probability score threshold comprises a decimal value between zero and one, inclusive.
10. The computer-implemented method of claim 1 , wherein activating the user device from the standby mode to perform the task comprises activating the user device from the standby mode to analyze additional digital representations of speech received after the digital representation of speech, the additional digital representations of speech corresponding to speech commands.
11. A user device:
data processing hardware; and
memory hardware in communication with the data processing hardware and storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:
receiving a digital representation of speech recorded by a microphone of the user device;
generating, using a first neural network executing on the user device, a probability score representing a likelihood that a specific keyword or key phrase is present in the digital representation of speech, the first neural network trained to identify the specific keyword or key phrase;
determining whether the probability score satisfies a probability score threshold; and
in response to determining that the probability score satisfies the probability score threshold:
generating, using a second neural network executing on the user device, a classification result confirming whether or not the specific keyword or key phrase is present in the digital representation of speech, the second neural network trained to identify the specific keyword or key phrase and having a higher accuracy than the first neural network; and
in response to the classification result confirming that the specific keyword or key phrase is present in the digital representation of speech, activating the user device from a standby mode to perform a task.
12. The user device of claim 11 , wherein the digital representation of speech comprises a stream of fixed-length audio frames.
13. The user device of claim 11 , wherein the first neural network executing on the user device requires less computation than the second neural network executing on the user device.
14. The user device of claim 11 , wherein:
the classification result comprises a posterior probability score representing the likelihood that the keyword or key phrase is present in the digital representation of speech; and
the classification result confirms that the specific keyword or key phrase is present in the digital representation of speech when the posterior probability score satisfies a posterior probability score threshold.
15. The user device of claim 11 , wherein:
the first neural network is trained to identify the specific keyword or key phrase using a first training set for a first quantity of iterations; and
the second neural network is trained to identify the specific keyword or key phrase using a second training set for a second quantity of iterations, the second quantity of iterations greater than the first quantity of iterations.
16. The user device of claim 15 , wherein the second training set and the first training set comprise the same training set.
17. The user device of claim 11 , wherein:
the first neural network comprises a first quantity of nodes; and
the second neural network comprises a second quantity of nodes, the second quantity of nodes greater than the first quantity of nodes.
18. The user device of claim 11 , wherein the first neural network comprises a higher false positive rate than the second neural network.
19. The user device of claim 11 , wherein the probability score threshold comprises a decimal value between zero and one.
20. The user device of claim 11 , wherein activating the user device from the standby mode to perform the task comprises activating the user device from the standby mode to analyze additional digital representations of speech received after the digital representation of speech, the additional digital representations of speech corresponding to speech commands.