63 histograms of Figure 58 while containing nearly as much information They are

63 histograms of figure 58 while containing nearly as

This preview shows page 73 - 76 out of 122 pages.

Image of page 73
histograms of Figure 5.8, while containing nearly as much information. They are far more informative than statistics such as standard deviation and bias, which are used frequently to evaluate performance. From Figure 5.10, it is easy to see that the data from the closest source, source 3, produced many more accurate estimates than the other two. Less than 10 percent of the estimates from source 3 had error greater than or equal to 0.5 samples (or 90 percent had error less than 0.5 samples) while about 50 percent of the others’ estimates had error greater than 4 samples. 5.2.3 RMS TDOA Error Rates with Speech Sources The experiment was run again using the array recordings of speech and the same triad array. As described in Section 3.2, there are three speech recordings in the conference-room data set, and the loudspeaker location during each was the same as it was during the Gaussian noise recordings. TDOA estimation was again performed using the same parameters: a Hanning window, 25-millisecond blocks, a 12.5-millisecond block advance, and a 0.1-sample TDOA resolution. The procedure was the same as the one applied to the Gaussian noise recordings, with one additional step. An SNR mask was used to discard any TDOAs that were produced by low-SNR speech blocks. The mask was derived using a 0.33 threshold and applied as explained in Section 3.5. Out of 399 blocks per recording, the mask passed 313 from source 1, 340 from source 2 and 297 from source 3. These speech recordings were processed using two GCC weighting functions. The first was the high-pass weighting function from Section 5.1.1, which was also applied to the Gaussian noise recordings. The second was a combination of this high-pass weighting function and the phase transform (PHAT) weighing function, which was presented in Section 4.1.2. RMS error rates were computed using the TDOA estimates produced by the speech recordings. These are shown in Figure 5.11. Also shown in this figure for comparison sake are the error rates for the corresponding Gaussian sources from Figure 5.10. Hence, for each source location, there are three error rates. The first two were derived from GCC with the high-pass weighing function, one for the Gaussian signals (labeled “Gaussian / GCC”) and one for the speech signals (labeled “Speech / GCC”). The third error rate in each plot was from GCC with the combined GCC-PHAT and high-pass weighing function for the speech signals (labeled “Speech / GCC-PHAT”). 64
Image of page 74
Figure 5.11 Error rates for the three source locations and two source signals at each location: Gaussian and speech. The speech sources were processed using both GCC and GCC-PHAT with SNR masks. In Figure 5.11, the Gaussian-source error rates serve as the baseline performance of GCC with the triad array in the conference room environment. These signals have properties that are nearly ideal and give GCC, in some sense, the “best you can do” results. The error rates reflect this, showing that the Gaussian sources consistently produce estimates that are more accurate than the speech sources. In general,
Image of page 75
Image of page 76

You've reached the end of your free preview.

Want to read all 122 pages?

  • One '14

What students are saying

  • Left Quote Icon

    As a current student on this bumpy collegiate pathway, I stumbled upon Course Hero, where I can find study resources for nearly all my courses, get online help from tutors 24/7, and even share my old projects, papers, and lecture notes with other students.

    Student Picture

    Kiran Temple University Fox School of Business ‘17, Course Hero Intern

  • Left Quote Icon

    I cannot even describe how much Course Hero helped me this summer. It’s truly become something I can always rely on and help me. In the end, I was not only able to survive summer classes, but I was able to thrive thanks to Course Hero.

    Student Picture

    Dana University of Pennsylvania ‘17, Course Hero Intern

  • Left Quote Icon

    The ability to access any university’s resources through Course Hero proved invaluable in my case. I was behind on Tulane coursework and actually used UCLA’s materials to help me move forward and get everything together on time.

    Student Picture

    Jill Tulane University ‘16, Course Hero Intern

Ask Expert Tutors You can ask You can ask ( soon) You can ask (will expire )
Answers in as fast as 15 minutes