### Citations

231 |
Acoustic characteristics of American English vowels.
- HILLENBRAND, GETTY, et al.
- 1995
(Show Context)
Citation Context ...st 5, 2014 DRAFT 10 30! 35! 40! 45! 50! 55! 60! 65! /ae/! /ah/! /eh/! /ei/! /er/! /ih/! /iy/! /oa/! /oo/! /uh/! /uw/! W eig ht edsC RL BsVo lum e! Fig. 5. The CRLB volume for each of the 11 vowels in =-=[12]-=-. difficulty of estimating specific vowels by calculating the full CRLB matrix for each vowel. We use the vowels from male speaker “m01” in the Hillenbrand data set [12]. As before, we downsample the ... |

119 | Applications of entropic spanning graphs. - Hero, Ma, et al. - 2002 |

99 | Geodesic entropic graphs for dimension and entropy estimation in manifold learning,” - Costa, Hero - 2004 |

91 |
On some asymptotic properties of maximumlikelihood estimates and related Bayes estimates,” Univ.
- Cam
- 1953
(Show Context)
Citation Context ...t of this work by ARO grant W911NF-11-1-0391 and NSF grant CCF-1217880. August 5, 2014 DRAFT ar X iv :1 40 8. 11 82 v1s[ sta t.C O]s6sA ugs20 14 2of the FIM [2], or develop super-efficient estimators =-=[3]-=-. The FIM can be computed as the covariance matrix of the gradient of the log likelihood function (the score function). However, in some cases the statistical model may not be known or the covariance ... |

91 |
Multivariate generalizations of the Wald-Wolfowitz and Smirnov two-sample tests.
- Friedman, Rafsky
- 1979
(Show Context)
Citation Context ...overlapping distributions and (b) separable distributions. p(x) = q(x). An underexplored, non-parametric estimator for this scenario is the multivariate runs test proposed by Friedman and Rafsky (FR) =-=[4]-=- based on constructing the the minimal spanning tree over Xp and Xq. This is a multivariate generalization of the Wald−Wolfowitz test for the two sample problem [5]. In the one-dimensional case, obser... |

79 |
On a test whether two samples are from the same population.
- WALD, WOLFOWITZ
- 1940
(Show Context)
Citation Context ...roposed by Friedman and Rafsky (FR) [4] based on constructing the the minimal spanning tree over Xp and Xq. This is a multivariate generalization of the Wald−Wolfowitz test for the two sample problem =-=[5]-=-. In the one-dimensional case, observations from both distributions are ranked in ascending order. Each observation is then replaced by a binary variable corresponding to the class to which it belongs... |

62 |
Information theory and statistics: A tutorial,” Foundations and Trends
- Csiszár, Shields
- 2004
(Show Context)
Citation Context ...n define the Henze-Penrose (HP) probability distance measure as DHP(p, q) = 1− 2A(p, q). It is easy to show that this divergence measure belongs to the class of f -divergences or Ali-Silvey distances =-=[7]-=-. Intuitively, an f -divergence is an average of the ratio of two distributions, weighted by some function f(t): Df (p, q) = ∫ f(p(x) q(x) )q(x)dx. Many common divergences used in statistical signal p... |

53 |
Empirical Bayes Estimation of the Multivariate Normal Covariance Matrix
- Haff
- 1980
(Show Context)
Citation Context ...mate Fθ using the SDP-based estimator in section 3 and we invert to calculate the CRLB, Cθ = F−1θ . Prior to inverting, we use the Bayesian-based diagonal loading procedure by Haff for regularization =-=[14]-=-. For each of the vowels in the analysis, we compute the volume of the CRLB matrix, weighted by a perceptual weighting filter giving more weight to bands that contain more signal energy. The weighting... |

36 |
Fedorov,Theory of Optimal Experiments,
- V
- 1972
(Show Context)
Citation Context ...thor gratefully acknowledges partial support of this work by ARO grant W911NF-11-1-0391 and NSF grant CCF-1217880. August 5, 2014 DRAFT ar X iv :1 40 8. 11 82 v1s[ sta t.C O]s6sA ugs20 14 2of the FIM =-=[2]-=-, or develop super-efficient estimators [3]. The FIM can be computed as the covariance matrix of the gradient of the log likelihood function (the score function). However, in some cases the statistica... |

35 |
Mimicking the human ear
- Loizou
- 1998
(Show Context)
Citation Context ...iculty identifying statistically significant differences between the true signal energy and no energy. This is consistent with the noisy and poor quality speech associated with real cochlear implants =-=[13]-=-. Performing this analysis for voiced and unvoiced speech segments independently reveals that the CRLB almost always dominates the signal energy for unvoiced; however, this is not the case for voiced ... |

22 |
On the multivariate runs test
- Henze, Penrose
- 1999
(Show Context)
Citation Context ...data set. We define C(Xp, Xq) as the number of edges of G(E,Xp∪Xq) connecting a data point from p to a data point from q. Henze and Penrose proved the following theorem related to this test statistic =-=[6]-=-: August 5, 2014 DRAFT 4Theorem 1. As Np →∞ and Nq →∞ in a linked manner such that NpNp+Nq → α, C(Xp, Xq) Np +Nq → 2α(1− α) ∫ p(x)q(x) αp(x) + (1− α)q(x)dx almost surely. In Fig. 5 and 5 we show two e... |

21 |
A contribution to the theory of statistical estimation
- Cramér
- 1946
(Show Context)
Citation Context ... information matrix (FIM) is an important quantity in signal processing and statistical estimation. It can be used to benchmark the performance of an estimator (via the Cramer-Rao Lower Bound (CRLB)) =-=[1]-=-, design an optimal experiment to maximize the trace or determinant V. Berisha is with the School of Electrical, Computer, and Energy Engineering and the Department of Speech and Hearing Science, Ariz... |

15 |
Amplitude mapping and phoneme recognition in cochlear implant listeners,’’ Ear Hear. 20, 60–74.1361Kong et al.: Combined acoustic and electric hearing
- Zeng, Galvin
- 1999
(Show Context)
Citation Context ...ech segments. This is consistent with results from the literature, where subjective tests show that individuals with cochlear implants have higher intelligibility rates for vowels than for consonants =-=[11]-=-. In addition to calculating the diagonal CRLB for the TIMIT database, we also compare the August 5, 2014 DRAFT 10 30! 35! 40! 45! 50! 55! 60! 65! /ae/! /ah/! /eh/! /ei/! /er/! /ih/! /iy/! /oa/! /oo/!... |

13 |
Information geometry of divergence functions,”
- Amari, Cichocki
- 2010
(Show Context)
Citation Context ...ector that is assumed to be small, e.g., ||u||2. Amari showed that any f -divergence induces a unique information monotonic Riemannian metric, given by the Fisher information matrix (see Theorem 5 in =-=[8]-=-), Fθ = Eθ[∇pθ∇pTθ ]. Using a Taylor expansion he showed that any f -divergence measure is related to the FIM through the asymptotic relation DHP(pθ, pθ+u) = α(1− α)uTFθu+ o(||u||2). (4) Combining thi... |

10 |
The recognition of vowels produced by men, women, boys, and girls by cochlear implant patients using a six-channel CIS processor, The
- Loizou, Dorman, et al.
(Show Context)
Citation Context ...ents with cochlear implants. In Fig. 5 we plot the weighted CRLB volume for August 5, 2014 DRAFT 11 the vowels in the Hillenbrand dataset. A similar behavioral study was conducted by Dorman et. al in =-=[17]-=-. Using the same dataset, the authors encode the vowels with a simulated CI and conduct an intelligibility assessment task where they ask participants to correctly identify the encoded vowels. The aut... |

4 |
An improved spectral subtraction method for speech enhancement using a perceptual weighting filter
- Vizireanu, Ciochina
(Show Context)
Citation Context ... with the components given by the signal energy at the location of the 16 frequencies associated with θ. Perceptual weighting of frequency bands by signal energy is common in many speech applications =-=[15]-=-, [16]. The volume of the weighted CRLB matrix is given by Vol = log(det(V DWV T)), where Cθ = V DV T. This estimate of the volume serves as a proxy for the uncertainty associated with estimating the ... |

1 |
Analysis-by-synthesis speech coding method with truncation of the impulse response of a perceptual weighting
- Mauc, Navarro
- 1999
(Show Context)
Citation Context ...the components given by the signal energy at the location of the 16 frequencies associated with θ. Perceptual weighting of frequency bands by signal energy is common in many speech applications [15], =-=[16]-=-. The volume of the weighted CRLB matrix is given by Vol = log(det(V DWV T)), where Cθ = V DV T. This estimate of the volume serves as a proxy for the uncertainty associated with estimating the spectr... |