CiteSeerX — Table search results

Results 1 - 10 of 13,465

Table 5: Recognition results on the car database. NLN: Noise level normalization, RA: Reference adaptation. German Car Navigation Database WER [%]

in Noise Level Normalization And Reference Adaptation For Robust Speech Recognition

by Florian Hilger, Hermann Ney, Lehrstuhl Fur Informatik Vi 2000

Cited by 4

Table 2. Recognition test results for the CarNavigation corpus with different normalization techniques: Cepstral mean normal- ization (CMN), cepstral variance normalization (CVN), filter bank mean normalization (FMN), quantile equalization (QE) in train- ing and test (QETT), histogram normalization (HN) with silence fraction treatment (HNSIL), feature space rotation (ROT), and vo- cal tract length normalization (VTN). Logarithm (log) and 10th root (root) were used to reduce the dynamic range of the filterbank channels.

in Feature Space Normalization In Adverse Acoustic Conditions

by Sirko Molau, Florian Hilger, Hermann Ney 2003

"... In PAGE 1: ... Furthermore it is known that normalizing the variance of cepstral coefficients (CVN) helps to improve recognition in adverse condi- tions. Recognition test results for these techniques are summarized in Table2... In PAGE 2: ... Whereas in the city traffic test set at least a third of the words are still recognized correctly, essentially nothing is correctly rec- ognized in the highway test set. Cepstral mean normalization ( Table2 , #1) has no impact on the office test set, since there is no channel mismatch. It more than halves the word error rate (WER) on the city data, but it is not sufficient for the highway data.... In PAGE 2: ... To ensure that the coefficients are positive, the 10th root was used for dynamic range reduction instead of the logarithm. It turned out, that replacing the logarithm by the 10th root alone reduces the error rate similar to cepstral variance normalization ( Table2 , #3). As it was shown in [4], quantile equalization requires as little as one second of data to estimate the transformation function reli- ably.... In PAGE 2: ... Furthermore it is possible to combine quantile equalization with mean normalization in a way that does not induce additional delay. Table2 gives recognition results for joint quantile and mean normalization (#4). The total delay is 500ms with a window length of 1s to estimate the quantiles and the mean.... In PAGE 2: ... Then quantile equalization is applied to match the distri- bution of each training utterance to the target distribution, before acoustic models are trained on the normalized data. Even though the training data of the CarNavigation corpus were recorded in a clean office environment, a significant reduction in word error rate under mismatch conditions of up to 17% relative was obtained by training data normalization ( Table2 , #5). 5.... In PAGE 2: ... The choice of the compression function is only of interest at startup when the target histogram is estimated, since this histogram determines the distribution of training and test data after normalization. Recognition test summarized in Table2 show that estimating the full histogram (#6) alone does not yield better results in mis- match conditions than quantile equalization (#5). However, the offline approach does allow for some further refinements leading to significant improvements in recognition performance.... In PAGE 3: ... The adapted target histogram is used for normaliza- tion as before. Recognition tests ( Table2 ) show that explicit si- lence fraction treatment (#7) reduces the word error rate by another 7% to 20% relative to baseline histogram normalization (#6). 6.... In PAGE 4: ... Larger reductions were reported for simple tasks or acoustic models, whereas the WER reduction fell typi- cally well below 10% relative for large vocabulary systems with advanced acoustic modeling trained on a large amount of data. On the CarNavigation corpus we achieved between 6% and 8% rela- tive reduction of WER ( Table2 , #9) compared to the best offline system including cepstral mean subtraction, histogram normaliza- tion with silence fraction treatment, and feature space rotation. 8.... ..."

Cited by 1

Table 5. Comparing navigation for pages with different statuses

in Comprehensive personalized information access in an educational digital library

by Peter Brusilovsky 2005

"... In PAGE 8: ... As mentioned above, the normalized access rate is computed by dividing the number of accesses to pages in this category by the total number of available pages in this category. Table5 shows how the normalized access rate is computed for each status. As the data shows, the chance of visiting an arbitrary resource in KSII is very low (close to 0) since there is a large number of resources in the system.... ..."

Cited by 3

Table 1. Recognition results on the isolated word car navigation database. baseline: MFCC front end with log compression and cepstral mean normalization, 10th: 10th root compression and mean normalization [4], QE: quan- tile equalization [4], QEF: quantile equalization with filter combination. Isolated Word Car Navigation Database SNR Word Error Rate [%]

in Combining Neighboring Filter Channels to Improve Quantile Based Histogram Equalization

by Florian Hilger, Hermann Ney, Lehrstuhl Fur Informatik Vi, Olivier Siohan, Frank K. Soong

Cited by 1

Table 1 summarizes the results of experiments conducted with InductoBeast during normal business hours with human and robot traffic. During the navigation portion of testing, tasks were invented based on a randomly generated sequence of desired goal positions. These positions were given in turn to InductoBeast, which then had to complete the navigation tasks.

in Automatic Mapping of Dynamic Office Environments

by Clayton Kunz, Thomas Willeke, Illah R. Nourbakhsh

"... In PAGE 7: ... Table1 : Empirical results for InductoBeast in two office environments. Terms: open area A path through a physical open area navigable length The distance in feet between two nodes contact Non-fatal touch between robot and environment collision A robot-environment touch requiring restart navigation success rate Percentage of random navigation runs in which the robot reached its destination navigation failure rate Percentage of random... ..."

Table 9. WAAS Navigation Functional State Mapping (Continued)

in A System for Integrated Reliability and Safety Analyses

by Peter Kostiuk, Gerald Shapiro, Dave Hanson, Stephan Kolitz, Frank Leong, Gene Rosch, Marc Coumeri, Peter Scheidler, Charles Bonesteel

"... In PAGE 7: ...able 8. WAAS Navigation Functional State Mapping.......................................................... 44 Table9 .... In PAGE 54: ... It should be emphasized that the numerical results should be taken as no- tional, since we were unable to validate them at this time. Table9 lists results for WAAS system state reliabilities for a location at 35-degree latitude with a 5-degree mask angle, geo-synchronous communication satellite with one global spare, master station with master computer only, and nominal uplink antenna. The WAAS system was modeled as four independent subsystems: GPS satellites, geo-synchronous communication satellites, master station, and uplink antenna.... In PAGE 54: ... The results have been normalized to account for numerical approximations of the PAWS routine. Table9 . WAAS Navigation Functional Reliability Results WAAS Navigation Reliability Fully operational Augmented GPS w/integrity 0.... In PAGE 55: ...odels, e.g., there is no Degraded Modes 1, 2 or 3 for the Receiver model. The reliability values are the probabilities shown in Table9 that are combined with the metrics generated by the simulation program. The WAAS subsystem can be defined with a variety of options.... ..."