Estimating pitch using symmetry characteristics
An estimate of a pitch of a signal may be computed by using correlations of frequency portions of a frequency representation of the signal. An initial pitch estimate may be obtained and frequency portions of the frequency representation may be identified using multiples of the initial pitch estimate. Correlations of the frequency portions may be computed, and a score for the initial pitch estimate may be determined using the correlations. A second pitch estimate may be determined using the first score, and the process may be repeated.
1. A computer-implemented method for estimating pitch in speech processing, the method comprising:
obtaining a first frame of a time representation of a signal;
obtaining a frequency representation of a first frame of the signal;
obtaining a first pitch estimate for the first frame of the signal;
identifying a first plurality of frequency portions of the frequency representation using the first pitch estimate, the first plurality of frequency portions comprising a first frequency portion and a second frequency portion;
computing a first plurality of correlations using the first plurality of frequency portions, the first plurality of correlations comprising a first correlation between the first frequency portion and the second frequency portion;
computing a first score for the first pitch estimate using the first plurality of correlations;
obtaining a second pitch estimate for the first frame of the signal;
identifying a second plurality of frequency portions of the frequency representation using the second pitch estimate, the second plurality of frequency portions comprising a third frequency portion and a fourth frequency portion;
computing a second plurality of correlations using the second plurality of frequency portions, the second plurality of correlations comprising a second correlation between the third frequency portion and the fourth frequency portion;
computing a second score for the second pitch estimate using the second plurality of correlations;
determining an updated pitch estimate using the first score and the second score;
computing amplitudes for a plurality of harmonics of the signal using at least the updated pitch estimate to describe a voice corresponding to human speech; 1 and
using the computed amplitudes to perform at least one of: speech recognition, speaker verification, speaker identification, signal reconstruction, word spotting, or noise reduction 2 .
2. The method of claim 1 , wherein the first plurality of correlations further comprises (i) a third correlation between the first frequency portion and a reversed version of the second frequency portion, and (ii) a fourth correlation between the first frequency portion and a reversed version of the first frequency portion.
3. The method of claim 1 , wherein the first plurality of frequency portions partitions the frequency representation.
4. The method of claim 1 , wherein computing the first score comprises computing a likelihood or a log likelihood of each correlation of the first plurality of correlations.
5. The method of claim 1 further comprising continuously updating the updated pitch estimate by performing a golden section search or a gradient descent.
6. The method of claim 1 , wherein each frequency portion of the first plurality of frequency portions is centered at a multiple of the first pitch estimate.
7. The method of claim 1 , further comprising normalizing each frequency portion of the first plurality of frequency portions before computing the first plurality of correlations.
8. A system for estimating features of a harmonic signal in speech processing the system comprising one or more computing devices comprising at least one processor and at least one memory, the one or more computing devices configured to:
obtaining a first frame of a time representation of a signal;
obtain a frequency representation of a first frame of the signal;
obtain a first pitch estimate for the first frame of the signal;
identify a first plurality of frequency portions of the frequency representation using the first pitch estimate, the first plurality of frequency portions comprising a first frequency portion and a second frequency portion;
compute a first plurality of correlations using the first plurality of frequency portions, the first plurality of correlations comprising a first correlation between the first frequency portion and the second frequency portion;
compute a first score for the first pitch estimate using, the first plurality of correlations;
obtain a second pitch estimate for the first frame of the signal;
identify a second plurality of frequency portions of the frequency representation using the second pitch estimate, the second plurality of frequency portions comprising a third frequency portion and a fourth frequency portion;
compute a second plurality of correlations using the second plurality of frequency portions, the second plurality of correlations comprising a second correlation between the third frequency portion and the fourth frequency portion;
compute a second score for the second pitch estimate using the second plurality of correlations;
determine an updated pitch estimate using the first score and the second score;
computing amplitudes for a plurality of harmonics of the signal using at least the updated pitch estimate to describe a voice corresponding to human speech; 3 and
using the computed amplitudes to perform at least one of: speech recognition, speaker verification, speaker identification, signal reconstruction, word spotting, or noise reduction 4 .
9. The system of claim 8 , wherein the first plurality of correlations further comprises (i) a third correlation between the first frequency portion and a reversed version of the second frequency portion, and (ii) a fourth correlation between the first frequency portion and a reversed version of the first frequency portion.
10. The system of claim 8 , wherein the first plurality of frequency portions partitions the frequency representation.
11. The system of claim 8 , wherein computing the first score comprises computing a Fisher transformation of each correlation of the first plurality of correlations.
12. The system of claim 8 , wherein each frequency portion of the first plurality of frequency portions is centered at a multiple of the first pitch estimate.
13. The system of claim 8 , wherein the one or more computing devices are further configured to normalize each frequency portion of the first plurality of frequency portions before computing the first plurality of correlations.
14. The system of claim 8 , wherein the one or more computing devices are further configured to:
continuously update the updated pitch estimate by performing a golden section search or a gradient descent.
15. One or more non-transitory computer-readable media comprising computer executable instructions that, when executed, cause at least one processor to perform actions in speech processing comprising:
obtaining a first frame of a time representation of a signal;
obtaining a frequency representation of a first frame of the signal;
obtaining a first pitch estimate for the first frame of the signal;
identifying first plurality of frequency portions of the frequency representation using the first pitch estimate, the first plurality of frequency portions comprising a first frequency portion and a second frequency portion;
computing a first plurality of correlations using the first plurality of frequency portions, the first plurality of correlations comprising a first correlation between the first frequency portion and the second frequency portion;
computing a first score for the first pitch estimate using the first plurality of correlations;
obtaining a second pitch estimate for the first frame of the signal;
identifying a second plurality of frequency portions of the frequency representation using the second pitch estimate, the second plurality of frequency portions comprising a third frequency portion and a fourth frequency portion;
computing a second plurality of correlations using the second plurality of frequency portions, the second plurality of correlations comprising a second correlation between the third frequency portion and the fourth frequency portion;
computing a second score for the second pitch estimate using the second plurality of correlations;
determining an updated pitch estimate using the first score and the second score;
computing amplitudes for a plurality of harmonics of the signal using at least the updated pitch estimate to describe a voice corresponding to human speech; 5 and
using the computed amplitudes to perform at least one of: speech recognition, speaker verification, speaker identification, signal reconstruction, word spotting, or noise reduction 6 .
16. The one or more non-transitory computer-readable media of claim 15 , wherein the first pitch estimate was computed using a plurality of peak-to-peak distances.
17. The one or more non-transitory computer-readable media of claim 15 , wherein the frequency representation was computed using an estimated fractional chirp rate.
18. The one or more non-transitory computer-readable media of claim 15 , wherein the first plurality of correlations further comprises (i) a third correlation between the first frequency portion and a reversed version of the second frequency portion, and (ii) a fourth correlation between the first frequency portion and a reversed version of the first frequency portion.
19. The one or more non-transitory computer-readable media of claim 15 , wherein the first plurality of correlations further comprises (i) a correlation between each pair of the first plurality of frequency portions, (ii) a correlation between each pair of the first plurality of frequency portions, wherein one of the pair has been reversed, and (iii) a correlation between each frequency portion and a reversed version of itself.