R. C. Rose and D. B. Paul, "A hidden Markov model based keyword recognition system," in Proc ICASSP 90, vol. 2.24, Apr. 1990, pp. 129--132.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....spotting consists in detecting a more or less important set of keywords from the speech stream. This process gives the exact time position of a keyword. Word spotting systems based on hidden Markov models are considered more efficient at modeling arbitrary speech than template based systems [20, 21]. Two main approaches are found in the literature. The most obvious is to use a large vocabulary continuous speech recognition system (LVCSR) to produce a word string. Then, search algorithms are applied for keyword detection in that string [22] This approach is considered as giving the best ....

Richard C. Rose and D. B. Paul, "A hidden Markov model based keyword recognition system," in Proceedings of ICASSP 90, 1990, pp. 129--132, Albuquerque, New Mexico, United States.

....attempt to employ HMMs (which play an important role in speech recognition at the present time) A first attempt by Sch uble and Glavitsch [21] had several drawbacks. The new approach presented in this paper is similar to a recently proposed method for spotting keywords in a speech recording [18]. It would seem that in the future, HMM matching will also become popular in fields other than speech recognition [7] 4] 9] 10] HMM matching may supersede dynamic programming techniques as used for instance in [25] because in contrat to the dynamic 319 programming approach, with its ....

....principle by Robertson [17] For the case where the image Y(Q, D) consists of sequences of different lengths, the T ta root in (18) provides a normalization such that short sequences do not have preference over long sequences. The same normalization is successfully used in keyword spotting systems [18], 24] This completes the presentation of a document retrieval method based on HMMs and its relationship to probabilistic retrieval. We conclude this section by showing that our document retrieval method based on HMMs encompasses a conventional retrieval method that is based on a well known ....

R. C. Rose and D. B. Paul. A Hidden Markov Model Based Keyword Recognition System. In International Conference on Acoustics, Speech, and Signal Processing, pages 129-132, 1990.

....method for evaluating the phone level performance, new features should be proposed and evaluated. Variations in the catch all model would lead to an interesting experiment. For example, instead of using a generic catch all model, various near miss [2] or boundary speci c anti models could be used [20, 25]. Some of the anti model research has been motivated by computational issues. Instead of having to use a generic model which is large, smaller computationally ecient normalizing models are used. Each boundary speci c anti model is made of all the models most like the boundary itself. It is ....

R. Rose and D. Paul. A hidden markov model based keyword recognition system. 1990. 90

....segment. 3 Filler based KWS Although various solutions have been proposed towards the direct optimization of (2) as, e.g. in [4, 11] most of the keyword spotting approaches today prefer to preserve the optimality and simplicity of Viterbi DP by modeling the complete input [6] and explicitly [7] or implicitly [3] modeling non keyword segments by using so called filler or garbage models as additional reference models. In this case, we assume that non keyword segments are modeled by extraneous garbage models states q G (and grammatical constraints ruling the possible keyword non keyword ....

Rose, R.C. and Paul, D.B., "A hidden Markov model based keyword recognition system," Proc. of ICASSP'90, pp. 129-132, 1990.

....for the OOV word. Since the OOV word is part of the vocabulary, the n gram grammar treats the OOV word just like any other word in the vocabulary. Augmenting the word recognizer with the generic word model shown in Figure 1 is somewhat similar to using filler (or garbage) models for word spotting [10, 11]. However, there are two key distinctions which differentiate our approach from using filler models for word spotting. First, the entire word vocabulary is used in the search, whereas the generic word is intended only to cover OOV words. The second distinction is that accurate sub word recognition ....

R. Rose and D. Paul, "A hidden Markov model based keyword recognition system," Proc. of ICASSP, Albuquerque, 1990.

....and channel noise. The generation of keyword hypotheses is often followed by a verification stage. In the absence of a language model, competition between two generative HMMs in a decoding is equivalent to the formation of a likelihood ratio. The use of explicit filler models is reported in [171, 210, 175, 172, 129, 123, 94, 172, 122, 173, 174]. One of the earliest descriptions is provided by Rohlicek et al. 171] who report the use of whole word keyword models and a single filler model created by combining the Gaussian emission probability distributions from all the keyword models. Wilpon et al. 210] report one of the first uses of ....

....et al. 123] report an investigation into various forms of filler models, including phonetic , syllabic and word level fillers. Syllabic fillers were found to provide the best performance in this study. An additional verification stage for hypothesised keywords was introduced by Rose Paul [175], in which the likelihood of an alternate model, distinct from the filler model, was used to form a ratio with the likelihood of the keyword model. The additional verification stage was found to reduce the number of false alarms given by the keyword spotter. Discriminative training techniques for ....

R. C. Rose and D. B. Paul. A hidden Markov model based keyword recognition system. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing, pages 129--132. IEEE, 1990.

....a word or phrase, for example find EUROSPEECH PROCEEDINGS , but in general may also be a non linguistic sound. Word spotting systems based on Hidden Markov Models (HMMs) are considered in this report. These have proved more successful at modelling arbitrary speech than template based systems [5], 9] For each keyword, an HMM is trained using statistical estimation techniques. Non keyword speech is modelled by one or more HMMs. Word spotting is performed by running a continuous speech recogniser with the keyword and non keyword models, and the recogniser outputs the sequence of HMMs ....

....beam search decoder consisting of a parallel network of keywords and fillers, with silence enforced at the start and end of each sentence (figure 1) Whole word continuous density mixture Gaussian distribution HMMs are used to model the keywords, trained on keyword speech. Following Rose and Paul [5] a set of 43 monophone 2 models are used as background filler models, trained on non keyword speech. FILLER M FILLER 1 KEYWORD N sil sil KEYWORD 1 KEYWORD 2 Filler network Keyword network Figure 1: Baseline null grammar word spotting network Recognition is performed using the token ....

[Article contains additional citation context not shown here]

Rose, R. C. and Paul, D. B. A Hidden Markov Model Based Keyword Recognition System. Proc ICASSP, S2.24, pp129-132, Albuquerque, April, 1990

....all negative. Note that the anti phone is not used during lexical access. Its only role is to serve as a form of normalization for the segment scoring. In this way, it has similarities with techniques being used in wordspotting, which compare acoustic likelihoods with those of filler models [16, 17, 19]. The framework holds whether or not the segmentation is done implicitly or explicitly, or whether the segmentation space is exhaustive, or restricted in some way. The recognition results reported here used a constrained network, since this is what we regularly use in our recognizer and allows us ....

R. Rose and D. Paul. A hidden Markov model based keyword recognition system. In Proc. ICASSP, pages 129--132, Albuquerque, NM, April 1990.

....all negative. Note that the anti phone is not used during lexical access. Its only role is to serve as a form of normalization for the segment scoring. In this way, it has similarities with techniques being used in wordspotting, which compare acoustic likelihoods with those of filler models [17, 18, 20]. The likelihood or odds ratio was also used by Cohen to use HMMs for segmenting speech [1] The independence assumption between X and Y made to enable efficient decoding is somewhat suspect since overlapping segments are likely correlated with each other. It would therfore be worth examining ....

R. Rose and D. Paul. A hidden Markov model based keyword recognition system. In Proc. ICASSP, pages 129--132, Albuquerque, NM, April 1990.

....will be aligned to the unrotated part of the sequence by the Viterbi algorithm. This alignment leads to a high probability P r( Oj) if the two objects have a similar shape. The method presented here has been animated by techniques applied to word spotting tasks in speech recognition (see e.g. [11]) Once the HMM original HMM Filler Model 1 Filler Model 2 Figure 4. Concatenated HMMs for rotation invariant modeling structure shown in Fig. 4 has been trained and built for every image in the database, sketches can be presented to the system and P r( Oj) can be used in order to score every ....

R. C. Rose and D. B. Paul. A Hidden Markov Model Based Keyword Recognition System. In Proc. IEEE Intern. Conference on Acoustics, Speech, and Signal Processing (ICASSP), pp. 129--132, Albuquerque, 1990.

....is to incorporate some form of filler or garbage model which is used to absorb OOV words and non speech artifacts. This approach has been effectively used in key word spotting for example, where the recognizer vocabulary primarily contains key words, so that the filler models are used extensively [11, 8]. In these applications, non key words absorbed by the filler model are of little subsequent interest. Our work differs from these applications in that we are very interested with accurately recovering the underlying sub word sequence of an OOV word for the purpose of ultimately recognizing the ....

R. Rose and D. Paul, "A Hidden Markov Model Based Keyword Recognition System," Proc. ICASSP, Albuquerque, 129--132, 1990.

....including supervised and unsupervised adaptation, recognition error rejection, out of vocabulary word detection, and keyword spotting. A method that has been popular for word based confidence modeling is the comparison of the score of the hypothesized word with the score of a filler model [1,2,3,4]. It was later demonstrated that the use of a large vocabulary speech recognizer improves confidence modeling [5] The approach taken in this paper is to extend this by also considering the scores of words which are commonly confused with the hypothesized word, thus detecting such confusions. The ....

Rose, R.C. and Paul, D.B. "A Hidden Markov Model Based Keyword Recognition System," ICASSP-90 1:129--132, 1990.

....Section 4. 3 Filler based KWS Although various solutions have been proposed towards the direct optimization of (2) as, e.g. in [4, 11] most of the keyword spotting approaches today prefer to preserve the optimality and simplicity of Viterbi DP by modeling the complete input [6] and explicitly [7] or implicitly [3] modeling non keyword segments by using so called filler or garbage models as additional reference models. In this case, we assume that non keyword segments are modeled by extraneous garbage models states q G (and grammatical constraints ruling the possible keyword non keyword ....

Rose, R.C. and Paul, D.B., "A hidden Markov model based keyword recognition system," Proc. of ICASSP'90, pp. 129-132, 1990.

.... principal recognition network and the alternative one. An example of the information obtained from this comparison is the likelihood score ratio. The use of the ratio of the two recognition scores is straightforward and commonly used in the keyword spotting and utterance veri cation techniques [1, 2, 3]. But recognition score is not the only useful information. If well compared, the resulting word strings (that represent the main product of recognition) may give some insight into the nature of the recognition process. This sort of information has not been used, so far, for con dence measuring or ....

R. C. Rose and D. B. Paul, \A Hidden Markov Model based keyword recognition system", in Proceedings of 1990 ICASSP, Albuquerque, April 1990, vol. I, pp. 129{ 132.

....measures (CM s) represent a feasible way to express which of the recognized sequences are likely to be correct and which can be disregarded as incorrect. A rather simple technique, that has shown remarkable results, to generate confidence measures is known as Likelihood score ratio (LSR) [10]. It is done by normalizing the likelihood score resulting from the Viterbi decoding process This research was supported by CONACyT and by CICYT under contract TIC98 0423 C06 01 by the likelihood score produced by an alternative recognition network. In our work, we add other information ....

....respectively. Due to its unconstrained (and inaccurate) nature, the purpose of the alternative network is to model the unrestricted signal probability, P ( X) This procedure tends to approximate Bayes law in posterior probability calculation. Because its simplicity and its high performance [10], we consider this feature as our baseline. 2.2. Sequence alignment score (SAS) Our second feature is what we call sequence alignment score (SAS) In the calculation of likelihood score ratio, the scores of both recognizers have been considered, but the decoded strings (the main product of ....

R. C. Rose and D. B. Paul. A Hidden Markov Model based keyword recognition system. In Proceedings of 1990 ICASSP, volume I, pages 129--132, Albuquerque, April 1990.

....the range of inputs allowable at that point in the interaction) The ability to reject out of domain utterances is essential for the design of user friendly interfaces. A number of rejection approaches have been suggested in the past for rejection of putative hits in keyword spotting (e.g. [1, 2, 3, 4, 5, 6, 7]) for detection of out of vocabulary words (e.g. 8] and for utterance rejection (e.g. 8, 9] Some of these systems use a filler model to match nonkeyword speech. A typical filler model is a set of contextindependent phonetic models. Also, some systems use antikeyword models. For example, ....

....and Section 4 presents conclusions and future directions. 2. PHONE BASED CONFIDENCE MEASURE Let PH = fPH1 ; PH2 ; PHN g be a Viterbi decoded sequence of phones for a spoken utterance. Let O = fO1 ; O2 ; OT g be the acoustic observation sequence for the utterance. Equivalently, O = fO b[1] ; O e[1] O b[2] O e[2] O b[N ] O e[N ] g, where b[i] and e[i] denote, respectively, the beginning and ending frames of the i th phone. Note that b[1] 1 and e[N ] T . Although our recognition system uses contextdependent phones, context independent phones are ....

[Article contains additional citation context not shown here]

R.C. Rose and D.B. Paul, "A Hidden Markov Model Based Keyword Recognition System," 1990 IEEE ICASSP, pp. 129-132, 1990.

....successful not only for large vocabulary recognition systems, but for keyword spotting systems where the location of only a few words or phrases is desired. Typically, this is done by training HMMs for both the desired keywords and a filler model that attempts to match everything not a keyword [5, 6, 7]. Such systems can be both accurate and computationally far less expensive than a large vocabulary recognition system, while being flexible enough to handle unconstrained real world speech [8] Large vocabulary recognition systems, in contrast, typically use a sub word approach: rather than ....

R. C. Rose and D. B. Paul. A hidden Markov model based keyword recognition system. In Proc. ICASSP 90, pages 129--132. IEEE, 1990.

....possible to determine the exact threshold corresponding to some given number of false alarms for a keyword, since it is obviously not known whether a putative keyword detection is correct or a false alarm. One of the best introductions to HMM based wordspotting is Rose and Paul s seminal paper [59]. In it, they described one of the first wordspotters to be based on the Viterbi recognition paradigm. Viterbi recognition, since it is based on the calculation of the most likely sequence of models for the unknown speech, guarantees CHAPTER 4. PREVIOUS WORK IN SPOKEN MESSAGE RETRIEVAL jonathan ....

....of interest for the forward backward wordspotter is the fast indexing of spoken messages such as voice mail. Rose also mentions that his wordspotter runs constantly, with partial traceback through the Viterbi path, so that putative word hits can be identified before the end of the unknown speech [59]. The delay between the utterance of the keyword and its hypothesis by the wordspotter is given as being of the order of a few seconds, which indicates that the wordspotter, once initialised by the loading of the models and network, runs in real time. Real time operation can be considered fast ....

[Article contains additional citation context not shown here]

R. C. Rose and D. B. Paul. A Hidden Markov Model based Keyword Recognition System. In Proc. Int. Conf. Acoust., Speech, Sig. Processing, pages 129--132, Albuquerque, 1990. IEEE.

....in the test data. Finally the lattice method was used, with all lattices being generated with degree 8. To allow for fair comparison between network based and lattice based wordspotters, the network based results were improved by rescoring all putative keywords to obtain a ratio score, as in [8]. The results are obtained as a standard NIST Figure of Merit averaged across 0 to 10 false alarms per keyword per hour and are shown in Table 1. It can be seen that the lattice wordspotter Wordspotter FOM Network Clustered VOCIND garb model 58.85 Network 80 word VOCDEP garb model 66.93 ....

R.C. Rose, D.B. Paul, A Hidden Markov Model Based Keyword Recognition System, Proc. Int. Conf. Acoust., Speech and Sig. Processing, May 1990, pp129-132.

....part of speech (background speech) and to score keywords. In [RRRG89] the non keyword part of speech is modeled by a background model consisting of segments of the keyword models. In addition, a causal posterior probability scoring method is used. The wordspotting systems presented in [RP90] and in [WB91] use a concatenation of phoneme models to represent the non keyword part of speech. In [RP90] the score reported for each keyword is a duration normalized likelihood whereas in [WB91] a posterior probability is used to compute keyword end points and two sets of backward ....

....is modeled by a background model consisting of segments of the keyword models. In addition, a causal posterior probability scoring method is used. The wordspotting systems presented in [RP90] and in [WB91] use a concatenation of phoneme models to represent the non keyword part of speech. In [RP90] the score reported for each keyword is a duration normalized likelihood whereas in [WB91] a posterior probability is used to compute keyword end points and two sets of backward probabilities are computed to detect keyword starting points. In both [RJN 93] and [Wei93] a large vocabulary ....

R. C. Rose and D. B. Paul. A Hidden Markov Model Based Keyword Recognition System. In International Conference on Acoustics, Speech, and Signal Processing, pages 129--132, 1990.

No context found.

R. C. Rose and D. B. Paul, "A hidden Markov model based keyword recognition system," in Proc ICASSP 90, vol. 2.24, Apr. 1990, pp. 129--132.

No context found.

R. Rose and D. Paul. A Hidden Markov Model based Keyword Recognition System. In IEEE Int. Conference on Acoustics, Speech, and Signal Processing (ICASSP), pages 129--132, Albuquerque, New Mexico, 1990.

No context found.

R.C. Rose and D.B. Paul, 'A hidden Markov model based keyword recognition system', ICASSP 90, pp 129132.

No context found.

R. C. Rose and D. B. Paul, \A hidden Markov model based keyword recognition system," in Proceedings of the International Conference on Acoustics, Speech, and Signal Processing, 1990, pp. 129-132.

No context found.

Rose, R. C. u. D. B. Paul. A Hidden Markov Model based keyword recognition system. IEEE ICASSP, pp. 129-132, 1990.

First 50 documents

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