List of Figures

1. Graphic interpretation of the most common clustering techniques
2. Example of a passive aperture response to different incoming signals
3. ICSI Speaker Diarization for Broadcast News blocks diagram
4. Acoustic models for speaker clustering
5. Speaker turn duration histograms
6. Overlap histograms in RT06s conference room meetings
7. Main blocks involved in the meetings speaker diarization system
8. single-channel speaker diarization for meetings block diagram
9. Speaker models initialization based on Gaussian splitting
10. Cross-validation EM training algorithm
11. Energy-based detector blocks diagram
12. Left, filter over $\tilde{e}[n]$. Decision of silence in red after the thresholding.
13. State machine used to apply time constraints.
14. Hybrid Speech/non-speech detector blocks diagram
15. Clusters initialization blocks diagram
16. Friends-and-enemies clusters initialization process
17. Cluster models with Minimum duration and modified probabilities
18. Possible Speaker clustering errors due to clusters purity problems
19. Speech-silence histogram for a full meeting
20. Observed assignment of frames to Gaussian mixtures
21. Evaluation of metric 1 on two clusters given their models
22. Speech/non-speech histograms for different possible model complexities
23. Linear microphone array with all microphones equidistant at distance d
24. Filter and sum algorithm blocks diagram
25. filter-and-sum implementation blocks diagram
26. Cross-correlation values histograms for RT06s AMI meetings
27. Filter and Sum double-Viterbi delays selection
28. Two-step TDOA Viterbi decoding example, step 1
29. Two-step TDOA Viterbi decoding example, step 1 for an individual channel
30. Two-step TDOA Viterbi decoding example, step 2
31. Multichannel delayed signal sum using a triangular window
32. Locations information contained in the TDOA values
33. Fusion of TDOA values and acoustic features within the speaker diarization module
34. First merge cluster-pair BIC values and histogram for acoustic and TDOA features
35. Acoustic weight evolution with the number of iterations for meeting CMU_20050912-0900
36. Energy-based system errors depending on its segment minimum duration
37. Model-based system errors depending on its segment minimum duration
38. Individual meetings DER vs. SNR vs. number of microphones in the RT06s system
39. Development set SNR modifying the percentage of noise threshold adjustment
40. Development set SNR values modifying the Viterbi transition prob. weights in the F&S algorithm
41. Development set SNR values modifying the number of N-best values used for TDOA selection
42. DER for the model complexity selection algorithm using different CCR values
43. DER for the initial number of clusters algorithm using different CCR values
44. DER for the combination of complexity selection + initial number of clusters using different CCR values
45. DER variation with the number of parallel models used in CV-EM training
46. DER variation with the number of friends used in the friends-and-enemies initialization
47. DER variation with the percentage of accepted frames and used Gaussians in frame purification
48. DER scores for the baseline system setting the relative weights by hand on development data
49. DER evolution with the weight computation iterations
50. DER evolution changing the initial feature stream weights
51. DER variation with the number of Gaussian mixtures initially assigned to the TDOA models
52. DER variation with the CCR parameter in the agglomerate system
53. DER variation with the number of friends in the agglomerate system
54. DER variation with the number of EM iterations of a standard EM-ML training algorithm
55. DER variation with the number CV-EM parallel models
56. DER variation with the frame % acceptance for frame purification algorithm
57. DER variation with the Gaussian % used in the frame purification algorithm
58. DER Break-down by meeting for the RT05s conference data
59. DER break-down by show for the RT05s lecture data
60. DER break-down by show for the RT06s conference data
61. DER break-down by show for the RT06s lecture data