Systems and methods for energy efficient and low power distributed automatic speech recognition on wearable devices

Patent metadata

Field	Value
Publication number	US-10373630-B2
Application number	US-201715475329-A
Country	US
Kind code	B2
Filing date	Mar 31, 2017
Priority date	Mar 31, 2017
Publication date	Aug 6, 2019
Grant date	Aug 6, 2019

Abstract

Methods, apparatus, systems and articles of manufacture are disclosed for distributed automatic speech recognition. An example apparatus includes a detector to process an input audio signal and identify a portion of the input audio signal including a sound to be evaluated, the sound to be evaluated organized into a plurality of audio features representing the sound. The example apparatus includes a quantizer to process the audio features using a quantization process to reduce the audio features to generate a reduced set of audio features for transmission. The example apparatus includes a transmitter to transmit the reduced set of audio features over a low-energy communication channel for processing.

First claim

What is claimed is:

1. An apparatus comprising:
a voice-activated detector;
a key phrase detector;
a quantizer; and
a transmitter,
the voice-activated detector to activate the key phrase detector when an input audio signal is detected,
the key phrase detector to, when activated, identify a portion of the input audio signal including a sound to be evaluated, the sound to be evaluated organized into a plurality of audio features representing the sound, the key phrase detector to activate the quantizer and the transmitter when the sound to be evaluated is identified in the input audio signal,
the quantizer to, when activated, reduce the audio features to generate a reduced set of audio features; and
the transmitter to, when activated, transmit the reduced set of audio features over a low-energy communication channel.

2. The apparatus of claim 1 , wherein the low-energy communication channel includes a Bluetooth Low Energy communication channel.

3. The apparatus of claim 1 , wherein the audio features include mel frequency cepstral coefficients (MFCCs).

4. The apparatus of claim 1 , wherein the quantizer is to reduce the audio features by reducing a number of the audio features.

5. The apparatus of claim 4 , wherein the quantizer is to reduce the number of the audio features from 23 audio features to 13 audio features.

6. The apparatus of claim 1 , wherein the quantizer is to reduce the audio features by reducing a bit size for each of the audio features.

7. The apparatus of claim 6 , wherein the quantizer is to reduce the bit size by applying an A-law algorithm to encode bits of each audio feature according to a compression parameter.

8. The apparatus of claim 6 , wherein the quantizer is to reduce the bit size by discarding one or more bits of each audio feature.

9. The apparatus of claim 8 , wherein the quantizer is to discard 7 least significant bits and one most significant bit of each audio feature.

10. The apparatus of claim 1 , wherein the apparatus includes a wearable device.

11. The apparatus of claim 1 , further including a feature extractor to extract the audio features from the sound identified by the key phrase detector.

12. The apparatus of claim 1 , further including a companion device to process the reduced set of audio features received from the transmitter.

13. The apparatus of claim 12 , wherein the companion device includes a dequantization engine and an automated speech recognition engine to process the reduced set of audio features to extract and evaluate the sound associated with the reduced set of audio features.

14. The apparatus of claim 12 , wherein the companion device includes at least one of a cellular phone or a gateway device.

15. The apparatus of claim 1 , wherein the sound to be evaluated is to include speech.

16. The apparatus of claim 1 , further including a feature extractor to extract the audio features from the portion identified by the key phrase detector.

17. The apparatus of claim 1 , further including a compressor to compress the reduced set of audio features for transmission by the transmitter.

18. The apparatus of claim 1 , further including a microphone to detect the input audio signal.

19. A method comprising:
activating a first process to be executed by a voice-activated detector when an input audio signal is detected;
identifying, by executing an instruction of the first process with the voice-activated detector, a portion of the input audio signal including a sound to be evaluated;
activating a second process to be executed by a key phrase detector when the sound to be evaluated is identified;
activating, by executing an instruction of the second process with the key phrase detector, a quantizer and a transmitter when the sound to be evaluated is identified in the input audio signal;
organizing, by executing an instruction of the second process with the key phrase detector, the sound to be evaluated into a plurality of audio features representing the sound;
reducing, by executing an instruction of a third process with the quantizer, the audio features to generate a reduced set of audio features; and
transmitting, using the transmitter, the reduced set of audio features over a low-energy communication channel.

20. The method of claim 19 , wherein the low-energy communication channel includes a Bluetooth Low Energy communication channel.

21. The method of claim 19 , wherein the audio features include mel frequency cepstral coefficients (MFCCs).

22. The method of claim 19 , wherein the reducing of the audio features includes reducing a number of audio features.

23. The method of claim 22 , wherein the reducing of the number of audio features includes reducing 23 audio features to 13 audio features.

24. The method of claim 19 , wherein the reducing of the audio features includes reducing a bit size for each of the audio features.

25. The method of claim 24 , wherein the reducing of the bit size includes applying an A-law algorithm to encode bits of each audio feature according to a compression parameter.

26. The method of claim 24 , wherein the reducing of the bit size includes discarding one or more bits of each audio feature.

27. The method of claim 19 , further including processing the reduced set of audio features using a companion device to extract and evaluate the sound associated with the reduced set of audio features.

28. A tangible computer readable storage medium comprising computer readable instructions which, when executed, cause a voice-activated detector implemented by a processor to at least:
in response to detection of an input audio signal:
activate a key phrase detector to:
identify a portion of the input audio signal including a sound to be evaluated; and
in response to identification of the sound to be evaluated:
activate a quantizer and a transmitter;
organize the sound to be evaluated into a plurality of audio features representing the sound;
instruct the quantizer to reduce the audio features based on a quantization process to generate a reduced set of audio features; and
instruct the transmitter to transmit the reduced set of audio features over a low-energy communication channel.

29. A wearable apparatus comprising:
a voice-activated detector to activate a key phrase detector, a quantizer, and a transmitter when an input audio signal is detected;
the key phrase detector which, when activated, is to identify a portion of the input audio signal including a sound to be evaluated, the sound to be evaluated organized into a plurality of audio features representing the sound;
the quantizer which, when activated, is to reduce the audio features to generate a reduced set of audio features; and
the transmitter which, when activated, is to transmit the reduced set of audio features over a low-energy communication channel to a companion device for dequantization of the reduced set of audio features to extract speech information from the reduced set of audio features.

Assignees

• Ravindran Binuraj K
• Tharappel Francis M
• Datta Prabhakar R
• Bocklet Tobias
• Muchlinski Maciej
• Dorau Tomasz
• Bauer Josef G
• Shah Saurin
• Stemmer Georg
• Intel Corp

Inventors

• Ravindran Binuraj K
• Tharappel Francis M
• Datta Prabhakar R
• Bocklet Tobias
• Muchlinski Maciej
• Dorau Tomasz
• Bauer Josef G
• Shah Saurin
• Stemmer Georg

Classifications

• G10L15/02 Primary

  Feature extraction for speech recognition; Selection of recognition unit · CPC title

• G10L19/032

  Quantisation or dequantisation of spectral components · CPC title

• G10L2015/088

  Word spotting · CPC title

• G10L15/30

  Distributed recognition, e.g. in client-server systems, for mobile phones or network applications · CPC title

• G10L15/08 Primary

  Speech classification or search · CPC title