Query endpointing based on lip detection
Systems and methods are described for improving endpoint detection of a voice query submitted by a user. In some implementations, a synchronized video data and audio data is received. A sequence of frames of the video data that includes images corresponding to lip movement on a face is determined. The audio data is endpointed based on first audio data that corresponds to a first frame of the sequence of frames and second audio data that corresponds to a last frame of the sequence of frames. A transcription of the endpointed audio data is generated by an automated speech recognizer. The generated transcription is then provided for output.
1. A computer-implemented method, comprising:
receiving video data and audio data that is synchronized with the video data;
determining that a sequence of video frames of the video data includes a representation of a face with moving lips;
providing the sequence of video frames that includes the representation of the face with moving lips as input to a model that is trained to determine whether lip movement is associated with speech or an action other than speech;
based on providing the sequence of video frames that includes the representation of the face with moving lips as input to the model that is trained to determine whether lip movement is associated with speech or an action other than speech, receiving, from the model, an indication determining that (i) a first portion of the sequence of video frames that includes the representation of the face with moving lips is associated with speech and (ii) a second portion of the sequence of video frames that includes the representation of the face with moving lips is associated with an action other than speech;
based on determining that (i) the first portion of the sequence of video frames that includes the representation of the face with moving lips is associated with speech and (ii) the second portion of the sequence of video frames that includes the representation of the face with moving lips is associated with an action other than speech, extracting a portion of the audio data that is synchronized with the first portion of the sequence of video frames that includes the representation of the face with moving lips;
generating, by an automated speech recognizer, a transcription of the extracted portion of the audio data by performing speech recognition on the extracted portion of the audio data that is synchronized with the sequence of video frames without performing speech recognition on the audio data other than the extracted portion of the audio data that is synchronized with the sequence of video frames; and
providing, for output, the generated transcription.
2. The method of claim 1 , wherein:
determining that the sequence of video frames of the video data includes the representation of a face with moving lips comprises:
identifying one or more feature statistics for the representation of the face with moving lips; and
the model that is trained to determine whether lip movement is associated with speech or an action other than speech is configured to:
determine whether the one or more identified feature statistics include a feature statistic for lip movement associated with speech.
3. The method of claim 1 , comprising:
determining that the video data includes user motion; and
in response to determining that the video data includes user motion, determining that the sequence of video frames of the video data includes the representation of the face with moving lips.
4. The method of claim 1 , wherein:
the video data and the audio data that is synchronized with the video data are received from a smartphone; and
the video data is captured by a front-facing camera of the smartphone.
5. The method of claim 1 , wherein the extracted portion of the audio data encodes a voice query spoken by a user.
6. The method of claim 5 , comprising:
in response to determining that the first portion of the sequence of video frames of the video data includes the representation of the face with moving lips is associated with speech, activating a personal assistant system to process the voice query spoken by the user.
7. The method of claim 1 , wherein determining that the sequence of video frames of the video data includes the representation of the face with moving lips comprises:
obtaining multiple subsets of consecutive video frames of the video data; and
processing each of the subsets of consecutive video frames of the video data using a deep neural network configured to:
receive each subset of consecutive video frames;
compute, for each subsets of consecutive video frames, a confidence score that represents a likelihood that the subset of consecutive video frames includes a representation of the face with moving lips; and
determine that each confidence score satisfies a confidence score threshold.
8. The method of claim 1 , wherein the model that is trained to determine whether lip movement is associated with speech or an action other than speech is configured to:
obtain multiple subsets of consecutive video frames of the sequence of video frames from the video data; and
process each of the subsets of consecutive video frames of the sequence of video frames using a deep neural network configured to:
receive each subset of consecutive video frames of the sequence of video frames;
compute, for each subset of consecutive video frames of the sequence of video frames, a confidence score that represents a likelihood that the subset of consecutive video frames includes a representation of the face with lip movement associated with speech; and
determine that each confidence score does not satisfy a confidence threshold.
9. The method of claim 1 , wherein extracting the portion of the audio data that is synchronized with the first portion of the sequence of video frames that includes the representation of the face with the moving lips comprises:
identifying first audio data synchronized with a first video frame of the first portion of the sequence of video frames of the video data;
identifying second audio data synchronized with the last video frame of the first portion of the sequence of video frames of the video data; and
truncating the audio data before the first audio data and after the second audio data.
10. 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 comprising:
receiving video data and audio data that is synchronized with the video data;
determining that a sequence of video frames of the video data includes a representation of a face with moving lips;
providing the sequence of video frames that includes the representation of the face with moving lips as input to a model that is trained to determine whether lip movement is associated with speech or an action other than speech;
based on providing the sequence of video frames that includes the representation of the face with moving lips as input to the model that is trained to determine whether lip movement is associated with speech or an action other than speech, receiving, from the model, an indication that (i) a first portion of the sequence of video frames that includes the representation of the face with moving lips is associated with speech and (ii) a second portion of the sequence of video frames that includes the representation of the face with moving lips is associated with an action other than speech;
based on determining that (i) the first portion of the sequence of video frames that includes the representation of the face with moving lips is associated with speech and (ii) the second portion of the sequence of video frames that includes the representation of the face with moving lips is associated with an action other than speech, extracting a portion of the audio data that is synchronized with the first portion of the sequence of video frames that includes the representation of the face with moving lips;
generating, by an automated speech recognizer, a transcription of the extracted portion of the audio data by performing speech recognition on the extracted portion of the audio data that is synchronized with the sequence of video frames without performing speech recognition on the audio data other than the extracted portion of the audio data that is synchronized with the sequence of video frames; and
providing, for output, the generated transcription.
11. The system of claim 10 , wherein:
determining that the sequence of video frames of the video data includes the representation of a face with moving lips comprises:
identifying one or more feature statistics for the representation of the face with moving lips; and
the model that is trained to determine whether lip movement is associated with speech or an action other than speech is configured to:
determine whether the one or more identified feature statistics include a feature statistic for lip movement associated with speech.
12. The system of claim 10 , comprising:
determining that the video data includes user motion; and
in response to determining that the video data includes user motion, determining that the sequence of video frames of the video data includes the representation of the face with moving lips.
13. The system of claim 10 , wherein:
the video data and the audio data that is synchronized with the video data are received from a smartphone; and
the video data is captured by a front-facing camera of the smartphone.
14. The system of claim 10 , wherein the extracted portion of the audio data encodes a voice query spoken by a user.
15. A non-transitory computer-readable storage device encoded with computer program instructions that, when executed by one or more computers, cause the one or more computers to perform operations comprising:
receiving video data and audio data that is synchronized with the video data;
determining that a sequence of video frames of the video data includes a representation of a face with moving lips;
providing the sequence of video frames that includes the representation of the face with moving lips as input to a model that is trained to determine whether lip movement is associated with speech or an action other than speech;
based on providing the sequence of video frames that includes the representation of the face with moving lips as input to the model that is trained to determine whether lip movement is associated with speech or an action other than speech, receiving, from the model, an indication that (i) a first portion of the sequence of video frames that includes the representation of the face with moving lips is associated with speech and (ii) a second portion of the sequence of video frames that includes the representation of the face with moving lips is associated with an action other than speech;
based on determining that (i) the first portion of the sequence of video frames that includes the representation of the face with moving lips is associated with speech and (ii) the second portion of the sequence of video frames that includes the representation of the face with moving lips is associated with an action other than speech, extracting a portion of the audio data that is synchronized with the first portion of the sequence of video frames that includes the representation of the face with moving lips;
generating, by an automated speech recognizer, a transcription of the extracted portion of the audio data by performing speech recognition on the extracted portion of the audio data that is synchronized with the sequence of video frames without performing speech recognition on the audio data other than the extracted portion of the audio data that is synchronized with the sequence of video frames; and
providing, for output, the generated transcription.
16. The device of claim 15 , wherein:
determining that the sequence of video frames of the video data includes the representation of a face with moving lips comprises:
identifying one or more feature statistics for the representation of the face with moving lips; and
the model that is trained to determine whether lip movement is associated with speech or an action other than speech is configured to:
determine whether the one or more identified feature statistics include a feature statistic for lip movement associated with speech.
17. The device of claim 15 , comprising:
determining that the video data includes user motion; and
in response to determining that the video data includes user motion, determining that the sequence of video frames of the video data includes the representation of the face with moving lips.
18. The device of claim 15 , wherein:
the video data and the audio data that is synchronized with the video data are received from a smartphone; and
the video data is captured by a front-facing camera of the smartphone.
19. The device of claim 15 , wherein the extracted portion of the audio data encodes a voice query spoken by a user.
20. The method of claim 1 , comprising:
training the model using (i) a group of first training sets of video frames, each first training set including a representation of a face with moving lips that is associated with speech, and (ii) a group of second training sets of video frames, each second training set including a representation of a face with moving lips that are associated with an action other than speech.
21. The method of claim 1 , wherein the second portion of the sequence of video frames includes a representation of the face with moving lips that is associated with eating.
22. The method of claim 1 , wherein the second portion of the sequence of video frames includes a representation of the face with moving lips that is associated with coughing.