S. Kamppari, "Word and phone level acoustic confidence scoring for speech understanding systems," 1999, master's thesis, MIT.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....decoding step, for example the lattice density [7, 16] N bests list [14] or language models [15] etc. Some other works use a post classifier to combine features such as likelihood and other statistics gathered from the decoding process (e.g. the number of letters in word, etc. into one measure [6, 8, 13]. By far, however, the most popular techniques are based on the building of a so called anti model or alternate model [1, 2, 4, 9] Such an anti model is used to normalize the likelihood of an unknown observation sequence O 1 = o 1 ; o 2 ; o T ) by computing a ratio between the joint ....

S. O. Kamppari and T. J. Hazen. Word and phone level acoustic confidence scoring. In International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Istanbul, Turkey, June 2000.

....example, sub word structure information of the recognition language would be very useful in providing linguistic support for unseen words, since such generic sub word structural constraints are not restricted to a specific recognition vocabulary. Another example is the use of confidence modeling [111, 82, 48] to improve the recognition robustness for poorly articulated speech. In this approach, the recognition result is evaluated by scoring a set of chosen confidence features against an established confidence model, and mis recognized words can be rejected according to the confidence score. ....

.... One possible way to incorporate such reliability information is through word and utterance level rejection [82] However, this approach generally provides confidence information after the recognition phase, and as such the confidence score is usually measured from a set of chosen features [48], most of which are obtained after the recognition is done. In contrast, we attempt to incorporate reliability information directly into the search phase in order to help the recognizer find the correct path. In this thesis, we also introduce our work on such dynamic phonetic reliability modeling ....

[Article contains additional citation context not shown here]

S. Kamppari and T. Hazen, "Word and phone level acoustic confidence scoring," in Proc. ICASSP'00, Istanbul, Turkey, 2000.

....6.2 Experimental Background In this section, we provide some background knowledge for the experiments described in this chapter, including the basics of the confidence scoring module, the speech data, and the labeling of the data. The details of the confidence scoring framework can be found in (Kamppari and Hazen 2000; Hazen et al. 2000a; Hazen et al. 2000b) 6.2.1 Experimental Framework The confidence scoring module developed by Hazen et al. is based on a Bayesian formulation. For each recognition hypothesis, a set of confidence measures are computed to form a confidence feature vector. The feature vector ....

Kamppari, S. O. and T. J. Hazen (2000). Word and phone level acoustic confidence scoring. In Proc. ICASSP'00, Istanbul, Turkey, pp. 1799--1802.

....data. One possible way to incorporate reliability information is through word and utterance level rejection [4] However, this approach generally provides confidence information after the recognition phase, and as such the confidence score is usually measured from a set of chosen features [2], most of which are obtained after the recognition is done. In contrast, this work attempts to incorporate reliability information directly into the search phase in order to help the recognizer find the correct path. In this paper, we introduce the notion of dynamic reliability scoring that ....

S. O. Kamppari and T. J. Hazen, "Word and phone level acoustic confidence scoring," in Proc. ICASSP'00,Istanbul, Turkey, 2000.

....a score which is zerocentered with respect to the log of p( x) allowing the scores to be consistent across different observations. In practice, the catch all model that is used is an approximation of the p( x) model that would result from the weighted summation of the p( xju) models over all u [7]. In this work, the individual phonetic scores are never used as independent confidence scores. However, they are used to help generate word and utterance level features. All references to acoustic scores in the remainder of this paper refer to the normalized acoustic scores described above. ....

....in the word graph is augmented with a score for its respective word. Before the implementation of word level confidence scores, a heuristic word scoring method was utilized which generated scores based on the number of N best hypotheses each word appeared in and the rank of those N best hypotheses [7]. In the new version of the system, each arc in the word graph is augmented with the word level confidence scores generated from the recognizer. The parser performs a beam search through the graph combining the word scores with trained linguistic probabilities to generate a total score for each ....

S. Kamppari, Word and Phone Level Acoustic Confidence Scoring for Speech Understanding Systems. Master's thesis, MIT, 1999.

....of processing to identify the pronunciation (and possibly the spelling) of the OOVword. In contrast, word spotters typically make no use of the output of the filler models. 3. WORD CONFIDENCE SCORING In our system, word confidence scores are computed as a postprocessing stage after recognition [7, 9, 12]. To obtain the confidence scores we begin by extracting a set of confidence measures for each word from the computations performed during the recognition process. In our system ten different confidence measures are computed. These include such measurements as the average normalized log likelihood ....

S. Kamppari and T. Hazen, "Word and phone level acoustic confidence scoring," Proc. of ICASSP, Istanbul, 2000.

....higher average raw acoustic scores in general. In our system, the acoustic model scores are normalized using the following expression: 1) In this expression is an acoustic observation, represents an acoustic model label, and is the normalization model [12]. In our case, is an approximation of and is estimated from the full set of all acoustic models across all languages (with each language receiving equal weighting) This normalization scheme converts the acoustic scores from absolute density scores to relative density scores, ....

S. Kamppari and T. Hazen, "Word and phone level acoustic confidence scoring," ICASSP, Istanbul, June, 2000.

....techniques focus on an examination of the scores produced by the recognizer s acoustic models at the phonetic level. Because the raw acoustic scores are usually not particularly useful as confidence measures when used by themselves [1] methods for normalizing these scores are typically employed [3, 8, 13]. In this work all of the acoustic scores produced at the phonetic level are normalized against a catch all model. The normalization of the acoustic score does not affect the outcome of the recognition search but does allow the score produced for each phone to act as a phonetic level confidence ....

S. Kamppari and T. Hazen, "Word and phone level acoustic confidence scoring," In Proc. of ICASSP, Istanbul, 2000.

....Gaussians in the next iteration. Each iteration reduces the number of Gaussian components by one. The process is continued until the estimated model is reduced enough for it to be computed efficiently during recognition. Details of the clustering algorithm and distance metric can be found in [5]. 3. EXPERIMENTS 3.1. System Description To evaluate the word confidence scoring techniques, the utterances used for the evaluation process were actual spontaneous queries collected over the telephone by the JUPITER weather information system [8] The word confidence scoring techniques are ....

S. Kamppari, Word and Phone Level Acoustic Confidence Scoring for Speech Understanding Systems. Master's thesis, MIT, 1999.

No context found.

S. Kamppari, "Word and phone level acoustic confidence scoring for speech understanding systems," 1999, master's thesis, MIT.

No context found.

Kamppari, S. O. and T. J. Hazen (2000). Word and phone level acoustic confidence scoring. In Proc. ICASSP'00, Istanbul, Turkey, pp. 1799--1802.

No context found.

S.O. Kamppari, T.J. Hazen, "Word and phone level acoustic confidence scoring". Proc. ICASSP, Istanbul. Turkey, 2000. pp III-1799, III-1802.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC