Method for scoring in an automatic speech recognition system
View Patent ↗A system and method for speech recognition is provided. Embodiments may include receiving an audio signal at a first deep neural network (“DNN”) associated with a computing device. Embodiments may further include receiving the audio signal at a second deep neural network (“DNN”) associated with a computing device, wherein the second deep neural network includes fewer parameters than the first deep neural network. Embodiments may also include determining whether to select an output from the first deep neural network or the second deep neural network and providing the selected output to a decoder with an overall objective of speeding up ASR.
1. A computer-implemented method comprising:
receiving an audio signal at a first deep neural network associated with a computing device;
receiving the audio signal at a second deep neural network associated with the computing device, wherein the second deep neural network includes fewer parameters than the first deep neural network;
determining one or more phonetic class conditional probabilities of the audio signal, wherein the one or more phonetic class conditional probabilities include at least one of silence, vowels, and consonants;
determining whether to select an output from one of the first deep neural network and the second deep neural network based upon, at least in part, the one or more phonetic class conditional probabilities of the audio signal; and
selecting the output from one of the first deep neural network and the second deep neural network based upon, at least in part, the one or more phonetic class conditional probabilities of the audio signal.
2. The method of claim 1 , further comprising:
terminating a computation associated with at least one of the first deep neural network and the second deep neural network.
3. The method of claim 1 , wherein the second deep neural network is configured to share one or more trained weights from one or more layers of the first deep neural network.
4. The method of claim 1 , wherein the second deep neural network includes fewer hidden layers than the first deep neural network.
5. The method of claim 1 , wherein the first deep neural network and the second deep neural network are associated with different processing threads.
6. The method of claim 1 , wherein the one or more phonetic class conditional probabilities further include speech.
7. The method of claim 1 , wherein determining whether to select the output is based upon, at least in part, an entropy associated with the output from one of the first deep neural network and the second deep neural network.
8. The method of claim 1 , further comprising:
training a classifier for model selection using a training module, wherein an input to the training module includes the output from the second deep neural network.
9. The method of claim 1 , wherein determining whether to select the output is based upon, at least in part, phonetic look-ahead information.
10. A system comprising:
one or more processors configured to receive an audio signal at a first deep neural network associated with a computing device, the one or more processors further configured to receive the audio signal at a second deep neural network associated with a computing device, wherein the second deep neural network includes fewer parameters than the first deep neural network, the one or more processors further configured to determine one or more phonetic class conditional probabilities of the audio signal, wherein the one or more phonetic class conditional probabilities include at least one of silence, vowels, and consonants, the one or more processors further configured to determine whether to select an output from one of the first deep neural network and the second deep neural network based upon, at least in part, the one or more phonetic class conditional probabilities of the audio signal, the one or more processors further configured to select the output from one of the first deep neural network and the second deep neural network based upon, at least in part, the one or more phonetic class conditional probabilities of the audio signal.
11. The system of claim 10 , wherein the one or more processors are further configured to terminate a computation associated with at least one of the first deep neural network and the second deep neural network.
12. The system of claim 10 , the second deep neural network is configured to share one or more trained weights from one or more layers of the first deep neural network.
13. The system of claim 10 , wherein the second deep neural network includes fewer hidden layers than the first deep neural network.
14. The system of claim 10 , wherein the first deep neural network and the second deep neural network are associated with different processing threads.
15. The system of claim 10 , wherein the one or more phonetic class conditional probabilities further include speech.
16. The system of claim 10 , wherein determining whether to select the output is based upon, at least in part, an entropy associated with the output from one of the first deep neural network and the second deep neural network.
17. The system of claim 10 , wherein the one or more processors are further configured to train a classifier for model selection using a training module, wherein an input to the training module includes the output from the second deep neural network.
18. The system of claim 10 , wherein determining whether to select the output is based upon, at least in part, phonetic look-ahead information.