PITCH DETECTION ALGORITHM BASED ON MULTIBAND PWVT OF TEAGER ENERGY OPERATOR

• US 20170287510A1
• Filed: 03/31/2017
• Published: 10/05/2017
• Est. Priority Date: 03/31/2016
• Status: Active Grant

• Current Assignee: Omnispeech LLC
• Original Assignee: Omnispeech LLC
• Inventors: KHANAGHA, Vahid
• Granted Patent: US 10,403,307 B2
• Time in Patent Office: Days
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 2015/223 Execution procedure of a spoken command
• CPC Class Codes: G10L 25/06 the extracted parameters being correlation coefficients
• CPC Class Codes: G10L 25/18 the extracted parameters being spectral information of each sub-band
• CPC Class Codes: G10L 25/90 Pitch determination of speech signals

Abstract

A pitch detection method. Such a pitch detection method may have M-PWVT-TEO algorithm to detect a pitch value from a speech signal, apply a partial auto-correlation to a current signal with the pitch value to compensate the delay of the pitch value. Also, the pitch detection method may apply a full auto-correlation to the speech signal where the pitch value is not detected to recover on-sets of the speech signal.

Claims

1. The method of a pitch detection comprising:

applying M-PWVT-TEO algorithm to a signal to detect a pitch value;

taking the pitch value as an initial guess; and

applying a partial auto-correlation to a current signal with the initial guess to generate a current pitch value.

3. The method of claim 1, wherein the partial auto-correlation is computed for a set of delay values that that represents a maximum 10 percent variation of the initial guess.

5. The method of claim 1, wherein the M-PWVT-TEO algorithm comprises:

sampling the signal to generate a first discrete time signal;

applying a Teager Energy Operator (TEO) to the first discrete time signal to generate a second discrete time signal;

high-pass filtering the first discrete time signal to generate a high-pass filtered discrete time signal;

applying a TEO to the high-pass filtered discrete time signal to generate a third discrete time signal;

generating a first complex valued signal from the first discrete time signal;

generating a second complex valued signal from the second discrete time signal;

generating a third complex valued signal from the third discrete time signal;

computing a Pseudo Weigner Ville Transformation (PWVT) on the first complex valued signal to generate a first spectral representation of the signal;

computing a PWVT on the second complex valued signal to generate a second spectral representation of the signal;

computing a PWVT on the third complex valued signal to generate a high-pass filtered spectral representation of the signal;

generating a combined spectral representation from the first spectral representation and the second spectral representation;

computing, to generate a pitch candidate, a harmonic summation on at least one of the combined spectral representation, the first spectral representation and high-pass filtered spectral representation; and

deciding the pitch candidate as the pitch value if the pitch candidate is larger than a threshold.

6. The method of claim 3, further comprising:

applying a Hilbert Transform to the first discrete time signal to generate the first complex valued signal;

applying a Hilbert Transform to the second discrete time signal to generate the second complex valued signal; and

applying a Hilbert Transform to the third discrete time signal to generate the third complex valued signal,

wherein each of the first complex valued signal, the second complex valued signal and the third complex valued signal is low-pass filtered, buffered and down-sampled.

7. The method of claim 3, wherein the PWVT is computed by Fourier Transform to represent a frequency content of the signal over time.

9. The method of claim 3, wherein the harmonic summation is computed on the combined spectral representation if a relative amount of a low-frequency energy of the signal is larger than a pre-determined threshold.

10. The method of claim 3, wherein the harmonic summation is computed within a human pitch frequency range.

11. The method of claim 3, wherein when deciding the pitch candidate as the pitch value, the pitch candidate is a frequency location of a maximum value of the harmonic summation.

12. The method of claim 3, wherein when deciding the pitch candidate as the pitch value, the threshold is fifty times of a median value of at least one of the combined spectral representation, the first spectral representation and the high-pass filtered spectral representation.

13. The method of claim 3, wherein when sampling the signal, the signal is sampled in each frame of the speech signal.

14. The method of claim 13, wherein when deciding the pitch candidate as the pitch value, the threshold is a smaller value between the median value and an average value of harmonic summation values taken over more than one hundred frames of the signal in which the pitch value is not detected.

15. The method of claim 3, wherein the harmonic summation is computed on the high-pass filtered spectral representation if a low frequency noise energy of the signal is larger than a remaining high frequency noise energy of the signal.

16. The method of claim 3, wherein the harmonic summation is computed on the high-pass filtered spectral representation if a first frequency noise energy of the signal is larger than 5 times of a second frequency noise energy of the signal, and the first frequency noise energy is a noise energy of a first 1 kHz and the second frequency noise energy is the noise energy of a remaining 3 kHz.

17. The method of claim 3, wherein when computing the harmonic summation, the Rainer Martin (RM) algorithm is applied to provide minimum noise estimates of a past 1 sec of the signal.

2. The method of claim 2, further comprising:

applying a full auto-correlation to the signal in which the pitch value is not detected to generate a maximum correlation value; and

deciding a corresponding delay value of the maximum correlation value as the current pitch value if the maximum correlation value is larger than an adaptive threshold.

4. The method of claim 2, wherein when deciding the corresponding delay value as the current pitch value, the signal comprises frames of a speech signal, the M-PWVT-TEO algorithm is applied to each frame of the speech signal, the adaptive threshold is an average of correlation values of the frames in which the pitch value is not detected, and the average is updated in each frame.

8. The method of claim, wherein a low-frequency part of the combined spectral representation are taken from the second spectral representation and a remaining part of the combined spectral representation are taken from the first spectral representation.