Stolcke, A., and Shriberg, E. (1996). Statistical language modeling for speech disfluencies. In Proceedings ICASSP, vol. 1, pp. 405-408.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....that words at the beginning of a new phrase correlate less strongly with the last word of the preceding phrase than words within the same phrase. A similar effect has also been found in the neighborhood of filled pauses [9] As a consequence, a language model for spontaneous speech is proposed in [10], where different types of disfluencies (filled pauses, repetitions, and deletions) are predicted, and probabilities of following words are estimated on the basis of the fluent word sequence that was supposedly intended by the speaker. This approach, however, did not have a significant impact on ....

....deletions) are predicted, and probabilities of following words are estimated on the basis of the fluent word sequence that was supposedly intended by the speaker. This approach, however, did not have a significant impact on the recognition accuracy. One of the reasons for this result is noted in [10]: phrase (or clause) boundaries grossly violate the assumptions of the proposed model, because filled pauses strongly correlate with boundaries of linguistic segments. Thus, cleaning up the surrounding words to remove the disfluency can be counterproductive. In our approach, phrase boundaries ....

Andreas Stolcke and Elizabeth Shriberg. Statistical Language Modeling for Speech Disfluencies. In Proc. Int. Conf. on Acoustics, Speech and Signal Processing, volume 1, pages 405--408, Atlanta, 1996.

.... and abridged repairs [6] consist solely of editing terms (i.e. they have no corrections) Speech based dialog systems often attempt to identify speech repairs in the speech recognition phase (prior to parsing) so that speech repairs will not disrupt the speech recognizer s language model ( 6] 7] [8]) In such a system, it is then tempting to remove conjectured reparanda (corrected material) and editing terms from the input prior to further processing. There are two issues that need to be addressed in such an approach, one pertaining to dialog interpretation and the other to parsing. First, ....

Stolcke, A. and Shriberg, E. 1996. Statistical language modeling for speech disfluencies. In Proceedings of the International Conference on Audio, Speech, and Signal Processing (ICASSP).

....plausible explanation of the variation in ASR performance is the difference in speaking style. Recognizers diminished performance on spontaneous speech can be attributed to many factors, such as differences in sentence structure or additional disfluencies that would affect the ASR language model [6,13]. One of the biggest influences, however, is the variation in pronunciations seen in spontaneous speech. We have observed [2] that an increase in errors made by ASR systems cor cu htk ibm limsi dragon bbn philips rwth sprach sri ogi fonix 0 5 10 15 20 25 30 ASR System by Site Percent ....

Stolcke, A. and Shriberg, E. 1996. Statistical language modeling for speech disfluencies. In IEEE ICASSP-96, pp. 405--409. Atlanta, GA.

....produced by the talker who is speaking while composing the message. Spontaneous speech is known to have variations in speaking rate, speech disfluencies (hesitations, restarts, incomplete words or fragments, repeated words) and rearranging of word sequences or incorrect syntactic structures[25]. Subsequent system modules must be able to deal with both the structures of spontaneous speech and recognition errors. By associating confidence scores with each hypothesized word the semantic analyzer and dialog modules can choose to ignore uncertain items, that could be misrecognitions. ....

A. Stolcke, E. Shriberg, "Statistical Language Modeling for Speech Disfluencies, ICASSP-96, Atlanta, GA, I, pp. 405-408, May 1996.

No context found.

A. Stolcke and E. Shriberg, "Statistical language modeling for speech disfluencies," in Proc. ICASSP, 1996.

No context found.

A. Stolcke and E. Shriberg, "Statistical language modeling for speech disfluencies," in Proc. ICASSP, 1996.

No context found.

A. Stolcke and E. Shriberg, "Statistical language modeling for speech disfluencies", in Proc. ICASSP, vol. 1, pp. 405--408, Atlanta, May 1996.

....Switchboard conversational speech in that it is far more template based. As a starting point to incorporate more syntactic information, we used a loosely coupled model. We trained a POS tagger using the Switchboard Treebank data [13] and used it to tag our training and testing data. Similar to [14], we maintained the identity of some cue words (e.g. filled pauses and discourse markers) Given the tag sequence and the hidden event tokens, we modeled the joint probability of the POS sequence P and the event sequence E. During testing we find the event sequence that maximizes P (EjP ) for the ....

A. Stolcke and E. Shriberg, "Statistical language modeling for speech disfluencies," in Proc. ICASSP, 1996.

.... likelihoods to condition the LM on other knowledge sources (e.g. prosody) 13] A special type of hidden event LM can model speech disfluencies by allowing the hidden events to modify the word history; for example, a word deletion event would erase one or more words to model a false start [14]. Skip language models In this LM, words in the history are probabilistically skipped, allowing more distant words to take their places. The skipping probabilities associated with each word are estimated using expectation maximization. HMMs of N grams This LM consists of a hidden Markov ....

A. Stolcke and E. Shriberg, "Statistical language modeling for speech disfluencies", in Proc. ICASSP, vol. 1, pp. 405--408, Atlanta, May 1996.

....for the work reported here comes from speech language modeling. Experiments at the 1995 Johns Hopkins Language Modeling Workshop showed that the quality of a language model (LM) can be improved if both training and test data are segmented linguistically, rather than acoustically [8] We showed in [10] and [9] that proper modeling of filled pauses requires knowledge of linguistic segment boundaries. We found for example that segment internal filled pauses condition the following words quite differently from segment initial filled pauses. Finally, recent efforts in languagemodeling for ....

....k#1 #p#w k j#s## p##s#j#s#w k#1 #p#w k j#s## A corresponding Viterbi algorithm is used to find the most likely sequence of S and NO S (i.e. a segmentation) for a given word string. This language model is a full implementation of the model approximated in [8] The hidden disfluency model of [10] has a similar structure. As indicated in the formulae above, we currently use at most two words of history in the local conditional probabilities p##j##. Longer N grams can be used if more state information is kept. The local N gram probabilities are estimated from the training data by using ....

A. Stolcke and E. Shriberg. Statistical language modeling for speech disfluencies. In Proceedings IEEE Conference on Acoustics, Speech and Signal Processing,volume I, pages 405-- 408, Atlanta, GA, May 1996.

....various spontaneous speech effects, such as disfluencies and syntactically incomplete constructions that might affect the dependency model in particular. For example, we might want to try to filter disfluencies prior to parsing. This requires a separate model for disfluencies, possible similar to [7, 31]. Finally, there is a fundamental concern about the applicability of an N best rescoring approach to a complex linguistically based model like ours. The best achievable error rate in our N best lists was 30 , making at least every third word incorrect on average. The language model therefore has ....

Andreas Stolcke and Elizabeth Shriberg. Statistical language modeling for speech disfluencies. In Proceedings of the IEEE Conference on Acoustics, Speech and Signal Processing, volume I, pages 405-- 408, Atlanta, May 1996.

....for the work reported here comes from speech language modeling. Experiments at the 1995 Johns Hopkins Language Modeling Workshop showed that the quality of a language model (LM) can be improved if both training and test data are segmented linguistically, rather than acoustically [8] We showed in [10] and [9] that proper modeling of filled pauses requires knowledge of linguistic segment boundaries. We found for example that segment internal filled pauses condition the following words quite differently from segment initial filled pauses. Finally, recent efforts in languagemodeling for ....

....k Gamma1 ) Theta p( s j s w k Gamma1 )p(w k j s ) A corresponding Viterbi algorithm is used to find the most likely sequence of S and NO S (i.e. a segmentation) for a given word string. This language model is a full implementation of the model approximated in [8] The hidden disfluency model of [10] has a similar structure. As indicated in the formulae above, we currently use at most two words of history in the local conditional probabilities p( Deltaj Delta) Longer N grams can be used if more state information is kept. The local N gram probabilities are estimated from the training data by ....

A. Stolcke and E. Shriberg. Statistical language modeling for speech disfluencies. In Proceedings IEEE Conference on Acoustics, Speech and Signal Processing, volume I, pages 405-- 408, Atlanta, GA, May 1996.

....NLU systems are trained to interpret fluent utterances. Recent studies suggest that disfluency detection is also relevant at other levels of speech processing. For example, work on statistical language modeling has shown that perplexity is reduced if disfluencies are removed from the Ngram context [12]. Additional analyses suggest that speakers hesitate before less predictable words; thus, transition probabilities should be dynamically adjusted in the vicinity of hesitations [9] Automatic detection of disfluencies could also benefit higher level modeling, for example, the automatic ....

....on a statistical language model (LM) The LM yields a joint probability P (W;D) where W is the word sequenceand D are the disfluency events. From this we can obtain another posterior probability estimate P (DjW ) P (D;W ) P (W ) The LM used was a disfluency N gram model of the type used in [12], and was trained on 1.4 million words of Switchboard transcripts, hand annotated for disfluencies by LDC [6] Finally, we want to combine the DT classifier basedonacoustic information with the LM classifier based on word information for a combined estimate. This can be done as follows: P (DjW;X) ....

A. Stolcke and E. Shriberg. Statistical language modeling for speech disfluencies. In Proc. ICASSP, vol. 1, pp. 405--408, Atlanta, 1996.

....communication and speech processing by machine. Although historically disfluencies have been viewed as noisy events, and have received relatively little attention, a more recent focus on spontaneous speech has directed increased interest to disfluencies in both theoretical and applied fields [1,4,6,8,9,10,11]. The goal of the present work is to illustrate that disfluencies are not noise but rather show systematic distributions in various dimensions. This paper summarizes results from a large descriptive study aimed at revealing and modeling trends in the distribution and form of disfluencies in ....

Stolcke, A. and Shriberg, E.E. "Statistical language modeling for speech disfluencies," Proc. ICASSP, 405-408, 1996.

....other studies suggest a correlation between the location of hesitations and the predictability of following words. If such a correlation exists, it has potential implications for language modeling in speech applications. In previous work we found that cleaning up disfluencies reduces perplexity [2]. However, if there is a correlation between hesitations and word predictability, we would lose information if we eliminate disfluencies completely. Thus, in this study we seek to determine whether such a correlation exists, but using a method based on corpus statistics instead of human judgments. ....

....word and the word history were found to contribute to this difference. We also found that FLUENT transitions in hesitant sentences were more likely than FLUENT transitions in fluent sentences to contain unseen N grams. For language modeling, we conclude that the simple cleanup model proposed in [2] needs to be extended. Specifically, the presence of a hesitation should increase the probability of unlikely following words, and decrease the probability of likely following words. Furthermore, the nonlocal effects found suggest that hesitations should modify the language model beyond the ....

Stolcke, A. and Shriberg, E.E. "Statistical language modeling for speech disfluencies," Proc. ICASSP, 405-408, 1996.

....speech can be reduced simply by resegmenting the speech at linguistic boundaries and using a language model based on the same segmentation. ffl Explicit modeling of spontaneous speech phenomena such as disfluencies also requires modeling of linguistic (as opposed to acoustic) segment boundaries [15]. Similarly, sophisticated LMs modeling syntactic structure typically assume complete sentences as their input [12] The following excerpt from the Switchboard corpus [2] illustrates the discrepancies between acoustic and linguistic segmentations. Linguistic segment boundaries are marked by s , ....

A. Stolcke and E. Shriberg. Statistical language modeling for speech disfluencies. In Proceedings of the IEEE Conference on Acoustics, Speech and Signal Processing, vol. 1, pp. 405--408, Atlanta, 1996.

No context found.

Stolcke, A., and Shriberg, E. (1996). Statistical language modeling for speech disfluencies. In Proceedings ICASSP, vol. 1, pp. 405-408.

No context found.

A. Stolcke and E. Shriberg. Statistical language modeling for speech disfluencies. In Proceedings of the International Conference of Acoustics, Speech, and Signal Processing, 1996.

No context found.

A. Stolcke and E. Shriberg. "Statistical Language Modeling for Speech Disfluency." volume 1, pages 405--408, 1996.

No context found.

Stolcke, A. and Shriberg, E. (1996). Statistical Language Modeling for Speech Disfluencies. In Proceedings of the IEEE International Conference on Acoustics Speech and Signal Processing, volume 1, pages 405--408.

No context found.

Stolcke, A. & Shriberg, E.E. (1996). Statistical language modeling for speech disfluencies. Proc. International Conference on Acoustics, Speech and Signal Processing, 405-408.

No context found.

Stolcke, A. and E. Shriberg. 1996b. Statistical language modeling for speech disfluencies. In Proceedings of the International Conference on Audio, Speech and Signal Processing (ICASSP),May.

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