1. Multiple Pitch Tracking for Blind Source Separation Using a Single Microphone Joseph Tabrikian Dept. of Electrical and Computer Engineering Ben-Gurion University of the Negev Workshop on: Speech Enhancement and Multichannel Audio Processing Technion 22.2.2007 BGU

2. Outline  Motivation  Single source pitch estimation and tracking  Multiple source pitch estimation and tracking  Experiments  Conclusion BGU

3. Motivation  Speech enhancement  Sensitivity of many audio processing algorithms to interference. For example: Automatic speech/speaker recognition Speech/music compression  Single microphone blind source separation (BSS)  Karaoke BGU

4. Single Source - Modeling  Voice frames - harmonic model: additive Gaussian noise  In matrix notation: BGU

5. Single Source – Pitch Tracking  Maximum Likelihood (ML) estimator:  Pitch tracking: The data vector at the m th frame: - first-order Markov process:  Maximum A-posteriori Probability (MAP) pitch tracking via the Viterbi algorithm. (Tabrikian-Dubnov-Dickalov 2004) BGU

6. Single Source - Voicing Decision  Unvoiced model Colored Gaussian noise model:  Voiced/unvoiced decision by the Generalized Likelihood Ratio Test (GLRT): BGU (Fisher-Tabrikian-Dubnov 2006)

7. Multiple Sources  ML estimator of from under the model: with unknown signal and unknown (Gaussian) noise covariance: BGU (Harmanci-Tabrikian-Krolik 2000)

8. Multiple Sources  Voiced model: v includes other interferences. is unknown.  Using J overlapping subframes of size L s (2K+1<J< L s ): jth column of : BGU

9. Multiple Sources  Pitch tracking: The data vector at the m th frame: - first-order Markov process  Maximum A-posteriori Probability (MAP) pitch tracking via the Viterbi algorithm BGU

10. Multiple Sources - Voicing Decision  Unvoiced model Colored Gaussian noise model:  Voiced/unvoiced decision by the GLRT: BGU (Fisher-Tabrikian-Dubnov 2007)

11. Multiple Source Models  Exact ML for the strongest voiced signal, and “ locally ML ” for other voiced signals BGU Likelihood function

12. Experiments – Single Source BGU

13. Experiments - Two Sources BGU

14. Experiments – Voicing Decision BGU

15. Experiments - – Voicing Decision BGU

16. Conclusions  ML pitch estimation for single and multiple sources have been developed under the harmonic model for voiced frames.  The derived likelihood functions under the two models allow implementation of the Viterbi algorithm for MAP pitch tracking.  The GLRT for voicing decision is derived under the two models.  Future work: development of multiple hypothesis tracking methods for single microphone BSS. Adaptive estimation of the number of harmonics BGU