Cross-modal sequence distillation
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a video speech recognition model having a plurality of model parameters on a set of unlabeled video-audio data and using a trained speech recognition model. During the training, the values of the parameters of the trained audio speech recognition model fixed are generally fixed and only the values of the video speech recognition model are adjusted. Once being trained, the video speech recognition model can be used to recognize speech from video when corresponding audio is not available.
1. A method of training a video speech recognition model having a plurality of model parameters, the method comprising:
obtaining (i) an audio sequence representing an utterance and (ii) a corresponding video sequence of the utterance being spoken;
processing the audio sequence using a trained audio speech recognition model to generate a respective teacher probability distribution over a vocabulary of possible audio speech recognition outputs for each of a first plurality of output time steps; and
training the video speech recognition model using the teacher probability distributions generated by the trained audio speech recognition model, comprising:
generating a plurality of target transcription sequences using the teacher probability distributions, the generating comprising, for each target transcription sequence:
sampling a respective speech recognition output for each of the output time steps from the teacher probability distribution for the output time step, and
generating the target transcription sequence using the sampled audio speech recognition outputs for the output time steps;
processing the corresponding video sequence using the video speech recognition model in accordance with current values of the model parameters to generate a respective student probability distribution over a vocabulary of possible video speech recognition outputs for each of a second plurality of output time steps; and
determining a respective update to the current values of the model parameters for each of the plurality of target transcription sequences by determining a gradient with respect to the model parameters of a loss between the target transcription sequence and the student probability distributions.
2. The method of claim 1 , wherein the loss is a CTC loss.
3. The method of claim 1 , wherein determining a respective update to the current values of the model parameters for each of the target transcription sequences by determining a gradient with respect to the model parameters of a loss between the target transcription sequence and the student probability distributions comprises:
determining a likelihood of the target transcription sequence according to the student probability distributions by marginalizing over transcription sequences that are equivalent to the target transcription sequence; and
determining a gradient of a negative log of the likelihood with respect to the model parameters.
4. The method of claim 1 , wherein the vocabulary of possible audio speech recognition outputs include a blank output and wherein generating the target transcription sequence using the sampled audio speech recognition outputs comprises:
removing any blank outputs and any adjacent duplicate outputs from the sampled audio speech recognition outputs.
5. The method of claim 1 , wherein the possible audio speech recognition outputs, the possible video speech recognition outputs, or both include phonemes.
6. The method of claim 1 , wherein the possible audio speech recognition outputs, the possible video speech recognition outputs, or both include characters in a natural language alphabet.
7. The method of claim 1 , wherein the possible audio speech recognition outputs, the possible video speech recognition outputs, or both include words.
8. The method of claim 1 , wherein the vocabulary of possible audio speech recognition outputs is different from the vocabulary of possible video speech recognition outputs.
9. The method of claim 1 , wherein the audio speech recognition model and the video speech recognition model are not temporally aligned and the first plurality of output time steps includes a different number of time steps than the second plurality of output time steps.
10. The method of claim 1 , wherein an output layer of the trained audio speech recognition model is a softmax layer, and wherein processing the audio sequence using a trained audio speech recognition model to generate a respective teacher probability distribution over a vocabulary of possible audio speech recognition outputs for each of a first plurality of output time steps comprises:
processing the audio sequence using the trained audio speech recognition model with a temperature parameter for the softmax layer set to a first value that is greater than one.
11. The method of claim 1 , wherein an output layer of the video speech recognition model is a softmax layer and wherein processing the corresponding video sequence using the video speech recognition model in accordance with current values of the model parameters to generate a respective student probability distribution over a vocabulary of possible video speech recognition outputs for each of a second plurality of output time steps comprises:
processing the video sequence using the video speech recognition model with a temperature parameter for the softmax layer set to a second value that is greater than one.
12. The method of claim 11 , wherein the first value is the same as the second value.
13. A system comprising:
one or more computers; and
one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations for training a video speech recognition model having a plurality of model parameters, the operations comprising:
obtaining (i) an audio sequence representing an utterance and (ii) a corresponding video sequence of the utterance being spoken;
processing the audio sequence using a trained audio speech recognition model to generate a respective teacher probability distribution over a vocabulary of possible audio speech recognition outputs for each of a first plurality of output time steps; and
training the video speech recognition model using the teacher probability distributions generated by the trained audio speech recognition model, comprising:
generating a plurality of target transcription sequences using the teacher probability distributions, the generating comprising, for each target transcription sequence:
sampling a respective speech recognition output for each of the output time steps from the teacher probability distribution for the output time step, and
generating the target transcription sequence using the sampled audio speech recognition outputs for the output time steps;
processing the corresponding video sequence using the video speech recognition model in accordance with current values of the model parameters to generate a respective student probability distribution over a vocabulary of possible video speech recognition outputs for each of a second plurality of output time steps; and
determining a respective update to the current values of the model parameters for each of the plurality of target transcription sequences by determining a gradient with respect to the model parameters of a loss between the target transcription sequence and the student probability distributions.
14. The system of claim 13 , wherein determining a respective update to the current values of the model parameters for each of the target transcription sequences by determining a gradient with respect to the model parameters of a loss between the target transcription sequence and the student probability distributions comprises:
determining a likelihood of the target transcription sequence according to the student probability distributions by marginalizing over transcription sequences that are equivalent to the target transcription sequence; and
determining a gradient of a negative log of the likelihood with respect to the model parameters.
15. The system of claim 13 , wherein the vocabulary of possible audio speech recognition outputs is different from the vocabulary of possible video speech recognition outputs.
16. The system of claim 13 , wherein the audio speech recognition model and the video speech recognition model are not temporally aligned and the first plurality of output time steps includes a different number of time steps than the second plurality of output time steps.
17. One or more non-transitory computer-readable storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations for training a video speech recognition model having a plurality of model parameters, the operations comprising:
obtaining (i) an audio sequence representing an utterance and (ii) a corresponding video sequence of the utterance being spoken;
processing the audio sequence using a trained audio speech recognition model to generate a respective teacher probability distribution over a vocabulary of possible audio speech recognition outputs for each of a first plurality of output time steps; and
training the video speech recognition model using the teacher probability distributions generated by the trained audio speech recognition model, comprising:
generating a plurality of target transcription sequences using the teacher probability distributions, the generating comprising, for each target transcription sequence:
sampling a respective speech recognition output for each of the output time steps from the teacher probability distribution for the output time step, and
generating the target transcription sequence using the sampled audio speech recognition outputs for the output time steps;
processing the corresponding video sequence using the video speech recognition model in accordance with current values of the model parameters to generate a respective student probability distribution over a vocabulary of possible video speech recognition outputs for each of a second plurality of output time steps; and
determining a respective update to the current values of the model parameters for each of the plurality of target transcription sequences by determining a gradient with respect to the model parameters of a loss between the target transcription sequence and the student probability distributions.
18. The non-transitory computer-readable storage media of claim 17 , wherein determining a respective update to the current values of the model parameters for each of the target transcription sequences by determining a gradient with respect to the model parameters of a loss between the target transcription sequence and the student probability distributions comprises:
determining a likelihood of the target transcription sequence according to the student probability distributions by marginalizing over transcription sequences that are equivalent to the target transcription sequence; and
determining a gradient of a negative log of the likelihood with respect to the model parameters.
19. The non-transitory computer-readable storage media of claim 17 , wherein the vocabulary of possible audio speech recognition outputs is different from the vocabulary of possible video speech recognition outputs.
20. The non-transitory computer-readable storage media of claim 17 , wherein the audio speech recognition model and the video speech recognition model are not temporally aligned and the first plurality of output time steps includes a different number of time steps than the second plurality of output time steps.