1. A Robust Algorithm for Pitch Tracking David Talkin Hsiao-Tsung Hung

2. Outline NCCF Algorithm outline Post-processing

3. Normalized cross-correlation function

4. 4 k

5. Problem Between the times of 198.625 and 198.7 the correlation peak at twice the correct period is stronger and more consistent than the “true” peak. NCCF 5

6. 1.The local maximum in Φ corresponding to the “true” F0 for voiced speech is usually the largest and it close to 1.0. 2.When multiple maximum in Φ exist and have values close to 1.0, the maximum corresponding to the shortest period is usually the correct choice. 3.True Φ maximum in temporally adjacent analysis frames are located usually at comparable lags, since F0 is a slowly-varying function of time. 4.The “true” F0 occasionally changes abruptly by doubling or halving 5.Voicing tends to change states with low frequency. 6.The largest non-zero-lag maximum in Φ for unvoiced speech is usually consider ably less than 1.0. 7.The short-time spectra of voiced and unvoiced speech frames are usually quite different. 8.Amplitude tends to increase at the onset of voicing and to decrease at offset. 6

7. Algorithm outline 1.Provide two versions of the sampled speech data.(16k,2k Hz) 2.Compute the NCCF of the low sample rate signal for all lags in the F0 range of interest(50~500Hz).Record the location of local maximum in the first-pass NCCF. 3.Compute the NCCF of the high sample rate signal only in the vicinity of promising peaks found in the first pass. Search again for the local maximum in this refined NCCF to obtain improved peak location and amplitude estimates. 4.Each peak retained from the high-resolution NCCF generates a candidate F0 for that frame. At each frame the hypothesis that the frame is unvoiced is also advanced. 5.DP is used to select the set of NCCF peaks or unvoiced hypotheses across all frames that best match the characteristics mentioned above. 7

8. Post-processing 8

9. 9

10. 10

11. Post-processing 11