Voice activity detection using vocal tract area information

• US 10,460,749 B1
• Filed: 06/28/2018
• Issued: 10/29/2019
• Est. Priority Date: 06/28/2018
• Status: Active Grant

• Current Assignee: Nuvoton Technology Corporation (Winbond Electronics Corporation)
• Original Assignee: Nuvoton Technology Corporation (Winbond Electronics Corporation)
• Inventors: Ru, Powen, Paiz, Alex
• Primary Examiner(s): Shah, Bharatkumar S
• Application Number:
  US16/021,724
• Time in Patent Office: 488 Days
• Field of Search: 704232
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 25/12 the extracted parameters being prediction coefficients
• CPC Class Codes: G10L 25/21 the extracted parameters being power information
• CPC Class Codes: G10L 25/30 using neural networks
• CPC Class Codes: G10L 25/78 Detection of presence or absence of voice signals switching of direction of transmission by voice frequency in two-way loud-speaking telephone systems H04M9/10

Abstract

A voice activity detection (VAD) system includes an input processing module configured to receive an acoustic signal, convert the acoustic signal into an analog signal, and subsequently, a digital signal; an energy-based detection module configured to receive one of the analog/digital signals and determine a sound activity decision; an area-function-based detection module configured to derive an area-related function from the digital signal and use a machine learning method to output an area-based decision according to the area related function; and a VAD decision module configured to make a final VAD decision based on the sound activity decision from the energy-based detection module and the area-based decision from the area-function-based detection module.

Claims

1. A voice activity detection (VAD) system, comprising:

a microphone interface circuit configured for coupling to a microphone to receive an acoustic signal and to convert the acoustic signal to an analog signal;

an analog-to-digital converter configured to receive the analog signal to generate a digital signal; and

a signal processing circuit configured to receive the digital signal and to determine if the digital signal represents a human voice, wherein the signal processing circuit comprises;

an acoustic-energy-based detection module configured to receive one of the analog signal or the digital signal and to provide a sound activity decision that indicates if the acoustic signal is in an audible energy range;

an area-function-based detection module configured to extract features of the acoustic signal from the digital signal based on area-related functions, and to use a machine-learning method to determine an area-based decision that indicates if the acoustic signal represents a human voice, wherein the machine-learning method comprises a plurality of coefficients trained by a plurality of labeled area-related functions; and

a voice activity detection (VAD) decision module configured to make a final VAD decision based on the sound activity decision from the acoustic-energy-based detection module and the area-based decision from the area-function-based detection module; and

a resource-limited device configured to receive the final VAD decision to change an operating mode of the resource-limited device.

2. The system of claim 1, wherein the area-related function comprises one of a plurality of log-area-ratios, a log area function, an area function, and a sagittal distance function.

3. The system of claim 1, wherein the area-function-based detection module is configured to perform:

filtering the digital signal with a pre-emphasis factor to obtain a pre-emphasized signal;

weighting a frame of the pre-emphasized signal to a windowed signal by a window function;

converting the windowed signal to a plurality of reflection coefficients;

converting the plurality of reflection coefficients to the area-related function;

feeding the area-related function to a trained classifier to identify onsets of voice; and

issuing the area-based decision.

4. The system of claim 3, wherein the trained classifier is trained offline by a neural network or a logistic regression.

5. A voice activity detection (VAD) system, comprising:

an input processing module configured to receive an acoustic signal via a microphone, the input processing module configured to convert the acoustic signal into an analog signal, and subsequently, a digital signal;

an energy-based detection module configured to receive one of the analog signal or the digital signal and determine a sound activity decision;

an area-function-based detection module configured to derive an area-related function from the digital signal and use a machine learning method to output an area-based decision according to the area-related function, wherein the machine learning method comprises a plurality of coefficients trained by a plurality of labeled area related functions;

and

a VAD decision module configured to make a final VAD decision based on the sound activity decision from the energy-based detection module and the area-based decision from the area-function-based detection module, wherein the final VAD decision is subsequently sent to a resource-limited device to change an operating mode of the resource-limited device.

6. The system of claim 5, wherein the energy-based detection module is a software module receiving the digital signal.

7. The system of claim 5, wherein the energy-based detection module is a digital hardware block receiving the digital signal.

8. The system of claim 5, wherein the energy-based detection module is an analog hardware block receiving the analog signal.

9. The system of claim 5, wherein the area-related function is a plurality of log-area-ratios.

10. The system of claim 5, wherein the area-related function comprises one of a plurality of log-area-ratios, a log area function, an area function, and a sagittal distance function.

11. The system of claim 5, wherein the sound activity decision is a soft decision value.

12. The system of claim 5, wherein the sound activity decision is a hard decision value.

13. The system of claim 5, wherein the area-function-based detection module is configured to perform the steps of:

filtering the digital signal with a pre-emphasis factor to obtain a pre-emphasized signal;

weighting a frame of the pre-emphasized signal to a windowed signal by a window function;

converting the windowed signal to a plurality of reflection coefficients;

converting the plurality of reflection coefficients to the area-related function;

feeding the area-related function to a trained classifier to identify onsets of voice; and

issuing the area-based decision.

14. The system of claim 13, wherein the pre-emphasis factor ranges from 0.5 to 0.99.

15. The system of claim 13, wherein a frame shift ranges from 1 millisecond to 20 milliseconds.

16. The system of claim 13, wherein the window function is one of Blackman, Blackman-Harris, Bohman, Chebyshev, Gaussian, Hamming, Hanning, Kaiser, Nuttall, Parzen, Taylor, and Tukey.

17. The system of claim 13, wherein the trained classifier is trained by a neural network offline.

18. The system of claim 13, wherein the trained classifier is trained by a logistic regression offline.

19. The system of claim 13, wherein the area-based decision is a soft decision value.

20. The system of claim 13, wherein the area-based decision is a hard decision value.

21. The system of claim 13, wherein the area-function-based detection module is configured to further generate a linear predictive error and include this error value to be a feature in the area-based decision.

22. The system of claim 5, further comprises a zero-crossing-based detection module configured to generate a second decision based on a zero crossing rate, wherein the VAD decision module includes the second decision in a final decision process.

23. The system of claim 22, wherein the second decision is a soft decision value.

24. The system of claim 22, wherein the second decision is a hard decision value.

25. The system of claim 5, wherein the resource-limited device is a low power device and the operating mode comprises an idle mode and a wake up mode.

26. The system of claim 5, wherein the resource-limited device is a voice storage device and the operating mode comprises an idle mode and a recording mode.

27. The system of claim 5, wherein the resource-limited device is a voice transmitting device and the operating mode comprises an idle mode and a transmitting mode.