Anchored speech detection and speech recognition

• US 10,373,612 B2
• Filed: 06/29/2016
• Issued: 08/06/2019
• Est. Priority Date: 03/21/2016
• Status: Active Grant

• Current Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Original Assignee: Amazon Technologies, Inc. (Amazon.com, Inc.)
• Inventors: Parthasarathi, Sree Hari Krishnan, Hoffmeister, Bjorn, King, Brian, Maas, Roland
• Primary Examiner(s): Baker, Matthew H
• Application Number:
  US15/196,228
• Publication Number:
  US 20170270919A1
• Time in Patent Office: 1,133 Days
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 15/20 Speech recognition techniques specially adapted for robustness in adverse environments, e.g. in noise, of stress induced speech G10L21/02 takes precedence
• CPC Class Codes: G10L 17/02 Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
• CPC Class Codes: G10L 17/06 Decision making techniques; Pattern matching strategies
• CPC Class Codes: G10L 17/18 Artificial neural networks; Connectionist approaches
• CPC Class Codes: G10L 2015/088 Word spotting
• CPC Class Codes: G10L 2025/783 based on threshold decision
• CPC Class Codes: G10L 25/87 Detection of discrete points within a voice signal

Abstract

A system configured to process speech commands may classify incoming audio as desired speech, undesired speech, or non-speech. Desired speech is speech that is from a same speaker as reference speech. The reference speech may be obtained from a configuration session or from a first portion of input speech that includes a wakeword. The reference speech may be encoded using a recurrent neural network (RNN) encoder to create a reference feature vector. The reference feature vector and incoming audio data may be processed by a trained neural network classifier to label the incoming audio data (for example, frame-by-frame) as to whether each frame is spoken by the same speaker as the reference speech. The labels may be passed to an automatic speech recognition (ASR) component which may allow the ASR component to focus its processing on the desired speech.

Claims

1. A computer implemented method for identifying speech from a desired speaker for automatic speech recognition (ASR), the method comprising:

receiving audio data corresponding to speech, the audio data comprising a plurality of audio frames;

processing the plurality of audio frames to determine a first plurality of feature vectors corresponding to a first portion of the audio data and a second plurality of feature vectors corresponding to a second portion of the audio data;

determining that the first plurality of feature vectors corresponds to a wakeword;

processing the first plurality of feature vectors with a recurrent neural network encoder to determine a reference feature vector corresponding to speech from a desired speaker;

processing the second plurality of feature vectors, and the reference feature vector, using a neural-network classifier to determine a first score corresponding to a first feature vector in the second plurality, the first score corresponding to a likelihood that the first feature vector corresponds to audio spoken by the desired speaker;

determining that the score is above a threshold;

creating an indication that the first feature vector corresponds to speech from the desired speaker;

determining a first weight corresponding to the first feature vector based on the first feature vector corresponding to speech from the desired speaker; and

performing ASR using the first weight and the first feature vector.

2. The computer-implemented method of claim 1, further comprising:

processing at least a portion of the plurality of audio frames to determine a third plurality of feature vectors corresponding to the second portion;

processing the third plurality of feature vectors and the reference feature vector using the neural-network classifier to determine a second score corresponding to a second feature vector in the third plurality, the second score corresponding to a likelihood that the second feature vector corresponds to audio spoken by the desired speaker;

determining that the second score is below the threshold;

creating a second indication that the second feature vector corresponds to speech from a different speaker as the wakeword; and

determining a second weight corresponding to the second feature vector based on the second feature vector corresponding to speech from a different speaker as the wakeword, wherein the second weight is less than the first weight.

3. The computer-implemented method of claim 1, further comprising:

identifying a first pair of feature vectors corresponding to audio frames positioned prior to a first audio frame corresponding to the first feature vector;

identifying a second pair of feature vectors corresponding to audio frames positioned after the first audio frame,

wherein processing the second plurality of feature vectors and the reference feature vector using the neural-network classifier further comprises processing the first pair of feature vectors, the first feature vector, and the second pair of feature vectors to determine the first score.

4. A computer implemented method comprising:

receiving first audio data as part of a first interaction with a device, the first audio data corresponding to first speech from a first speaker;

determining, using the first audio data and a recurrent neural network, a reference feature vector;

receiving second audio data as part of a second interaction with the device;

determining, using the reference feature vector and a trained model, that a first portion of the second audio data corresponds to second speech from a second speaker;

determining, using the reference feature vector and the trained model, that a second portion of the second audio data corresponds to third speech from the first speaker; and

based at least in part on determining that the first portion of the second audio data corresponds to the second speaker, executing a command corresponding to the second portion of the second audio data.

6. The computer-implemented method of claim 4, further comprising:

determining a first score corresponding to whether the reference feature vector corresponds to the third speech from the first speaker.

7. The computer implemented method of claim 6, further comprising:

determining a first weight corresponding to the reference feature vector based on the first score; and

performing speech recognition using the first weight and the reference feature vector.

8. The computer implemented method of claim 7, further comprising:

determining a second score corresponding to whether the reference feature vector corresponds to the second speech from the second speaker, wherein the second score is lower that the first score; and

determining a second weight corresponding to the second score, wherein the second weight is less than the first weight.

9. The computer-implemented method of claim 4, wherein the recurrent neural network is configured to input a plurality of feature vectors, and wherein determining the reference feature vector is based at least in part on each of the plurality of feature vectors.

16. The computer implemented method of claim 4, further comprising:

determining that the first audio data corresponds to a wakeword.

5. A computer implemented method comprising:

receiving audio data as part of an interaction with a device;

determining that a first portion of the audio data represents a wakeword spoken by a first speaker;

based at least in part on determining that the first portion of the audio data represents the wakeword, processing the first portion of the audio data to determine a reference feature vector;

determining, using the reference feature vector and a trained model, that a second portion of the audio data represents speech of a second speaker;

determining, using the reference feature vector and the trained model, that a third portion of the audio data represents speech of the first speaker; and

based at least in part on determining that the second portion of the audio data represents the speech of the second speaker, executing a command corresponding to the third portion of the audio data.

10. A computing system comprising:

at least one processor; and

at least one memory including instructions that, when executed by the at least one processor, cause the computing system to;

receive first audio data as part of a first interaction with a device, the first audio data corresponding to first speech from a first speaker;

determine, using the first audio data and a recurrent neural network, a reference feature vector;

receiving second audio data as part of a second interaction with the device;

determine, using the reference feature vector and a trained model, that a first portion of the first audio data corresponds to second speech from a second speaker;

determine, using the reference feature vector and the trained model, that a second portion of the second audio data corresponds to third speech from the first speaker; and

based at least in part on determining that the first portion of the second audio data corresponds to the second speaker, execute a command corresponding to the second portion of the second audio data.

12. The computing system of claim 10, wherein the at least one memory further includes instructions that further configure the system to:

determine a first score corresponding to whether the reference feature vector corresponds to the third speech from the first speaker.

13. The computing system of claim 12, wherein the at least one memory further includes instructions that further configure the system to:

determine a first weight corresponding to the reference feature vector based on the first score; and

perform speech recognition using the first weight and the reference feature vector.

14. The computing system of claim 13, wherein the at least one memory further includes instructions that further configure the system to:

determine a second score corresponding to whether the reference feature vector corresponds to the second speech from the second speaker, wherein the second score is lower that the first score; and

determine a second weight corresponding to the second score, wherein the second weight is less than the first weight.

15. The computing system of claim 10, wherein the recurrent neural network is configured to input a plurality of feature vectors, and wherein the determination of the reference feature vector is based at least in part on each of the plurality of the plurality of feature vectors.

17. The computing system of claim 10, wherein the at least one memory further includes instructions that further configure the system to:

determine that the first audio data corresponds to a wakeword.

11. A computing system comprising:

at least one processor; and

at least one memory including instructions that, when executed by the at least one processor, cause the system to;

receive audio data as part of an interaction with a device;

determine that a first portion of the audio data represents a wakeword spoken by a first speaker;

based at least in part on determining that the first portion of the audio data represents the wakeword, process the first portion of the audio data to determine a reference feature vector;

determine, using the reference feature vector and a trained model, that a second portion of the audio data represents speech of a second speaker;

determine, using the reference feature vector and the trained model, that a third portion of the audio data represents speech of the first speaker; and

based at least in part on determining that the second portion of the audio data represents the speech of the second speaker, execute a command corresponding to the third portion of the audio data.