I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. ICSLP, 401--404, Beijing, China, 2000.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....sections detail how to eliminate vocabulary items the recognizer finds little use for, and how to detect and resolve competition between similar items. Extracting OOV phone sequences Recognizer is that developed by the SLS group at MIT [8] The recognizer used the OOV model developed by Bazzi in [3]. This model can match an arbitrary sequence of phones, and has a phone bigram to capture phonotactic constraints. The OOV model is placed in parallel with the models for the words in the vocabulary. A cost parameter can control how much the OOV model is used at the expense of the in vocabulary ....

....how much the OOV model is used at the expense of the in vocabulary models. This value was fixed at zero throughout the experiments described in this paper, since it was more convenient to control usage at the level of the language model. The bigram used in this project is exactly the one used in [3], with no training for the particular domain. Recovering phonemic representations It is useful to convert the extracted phone sequences to phonemes if they are to be added as baseforms in the lexicon. Although the sequences could be kept in their original form by creating a dummy set of units for ....

I. Bazzi and J.R. Glass. Modeling out-of-vocabulary words for robust speech recognition. In Proc. 6th International Conference on Spoken Language Processing, Beijing, China, October 2000.

....the worth of the new vocabulary entries. The following sections detail how to eliminate vocabulary items the recognizer finds little use for, and how to detect and resolve competition between similar items. 2. 1 Extracting OOV phone sequences The recognizer used the OOV model described in [1] , contributed by Issam. This model can match an arbitrary sequence of phones, and has a phone bigram to capture phonotactic constraints. The OOV model is placed in parallel with the models for the words in the vocabulary. A cost parameter can control how much the OOV model is used at the expense ....

....how much the OOV model is used at the expense of the invocabulary models. This value was fixed at zero throughout the experiments described in this paper, since it was more convenient to control usage at the level of the language model. The bigram used in this project is exactly the one used in [1], with no training for the particular domain. 2.2 Recovering phonemic representations It is useful to convert the extracted phone sequences to phonemes if they are to be added as baseforms in the lexicon. Although the sequences could be kept in their original form by creating a dummy set of ....

Bazzi, J.R. Glass, Modeling Out-of-Vocabulary Words for Robust Speech Recognition , Proc. 6th International Conference on Spoken Language Processing, Beijing, China October 2000.

....feature (third last row) where the boundary is visually distinct to either side of the edge. This is followed by corner like features and many thousands of variations on the themes already seen. recognized words are explicitly represented using a phonebased OOV (out of vocabulary) model following [Bazzi and Glass, 2000] . The recognizer is then run on a large set of (untranscribed) acoustic data. The phonetic and word level outputs of the recognizer are compared so that occurrences of OOV fragments can be assigned a phonetic transcription. A randomly cropped subset of these are tentatively entered into the ....

I. Bazzi and J.R. Glass. Modeling out-of-vocabulary words for robust speech recognition. In Proc. 6th International Conference on Spoken Language Processing, Beijing, China, October 2000.

....vocabulary items the recognizer finds little use for, and how to detect and resolve competition between similar items. Extracting OOV phone sequences We use the speech recognizer system developed by the SLS group at MIT [8] The recognizer is augmented with the OOV model developed by Bazzi in [2]. This model can match an arbitrary sequence of phones, and has a phone bigram to capture phonotactic constraints. The OOV model is placed in parallel with the models for the words in the vocabulary. A cost parameter can control how much the OOV model is used at the expense of the in vocabulary ....

....how much the OOV model is used at the expense of the in vocabulary models. This value was fixed at zero throughout the experiments described in this paper, since it was more convenient to control usage at the level of the language model. The bigram used in this project is exactly the one used in [2], with no training for the particular domain. Phone sequences are translated to phonemes, then inserted as new entries in the recognizer s lexicon. Dealing with rarely used additions If a phoneme sequence introduced into the vocabulary is actually a common sound sequence in the acoustic data, ....

I. Bazzi and J.R. Glass. Modeling out-of-vocabulary words for robust speech recognition. In Proc. 6th International Conference on Spoken Language Processing, Beijing, China, October 2000.

....and square brackets are used for phones. 26 the in vocabulary words and build more flexible speech recognizers, sub lexical models can be built using solely statistical knowledge. For example, a general phone n gram model can be used to model both in vocabulary and out of vocabulary words [10]. With large amounts of training data, statistical models can capture the underlying sub lexical phonological knowledge by learning the probabilities of di#erent phone connections. However, since the n gram model encodes the linguistic knowledge implicitly by probability, it is hard to analyze the ....

....used in a standard keyword spotting system [90, 91] The FST has an in vocabulary branch that defines the phoneme network for all known words in the vocabulary, and a filler branch that accepts arbitrary phoneme sequences for modeling unseen words. The same topology was proposed by Bazzi and Glass [10] and studied in detail. The two FST branches are weighted by # and 1 #, and the latter controls the penalty of entering the 69 . ihr b aa ao s t en z ow Figure 3 2: FST topology for the phoneme network model. Each path represents a legitimate phoneme sequence ....

[Article contains additional citation context not shown here]

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition, " in Proc. ICSLP'00, Beijing, China, 2000.

....models. We have implemented an ANGIE [2] morpho phonemic model and a novel hybrid model which combines the ANGIE model with a lexicon based phoneme network model by constructing an FST with an in vocabulary branch and an ANGIE OOV branch. The same topology for the hybrid model was used in [4] except that the OOV branch is now modeled by ANGIE morpho phonemic rules. We have compared the ANGIE based models with some other models including a simple lexicon based phoneme network model, a phoneme network model with fillers and a phoneme n gram model. We also demonstrated the feasibility ....

....the recognizer performs significantly worse when OOV words are included. In order to allow OOV words in the recognizer, one approach is to use phoneme fillers to model and detect the OOV and partial words, with unique filler path for each phoneme. This is similar to another model for OOV words [4], in which a bigram model is used in the OOV branch. This model accepts any arbitrary phoneme sequence through fillers, which are also used in a standard keyword spotting system. The operation point (the false alarm and OOV detection rate) can be controlled by a penalty for detecting OOV words. It ....

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. of ICSLP, Beijing, 2000.

....occurred. As has been done previously in many other speech recognition systems, we have added a collection of non speech models to the recognizer to address this problem [4] To add non speech models to our system, we utilize an FST approach first developed for modeling out of vocabulary words [5]. We define a set of acoustic models and a network topology for each non speech type. Each network uses a fully connected topology allowing the noise to be represented by any sequence of the acoustic models used for that noise. It is possible to constrain the sequence of acoustic models used with ....

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," ICSLP, Beijing, October, 2000.

No context found.

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. ICSLP, 401--404, Beijing, China, 2000.

No context found.

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. ICSLP, 401--404, Beijing, China, 2000.

No context found.

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. ICSLP, 401--404, Beijing, China, 2000.

No context found.

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. ICSLP, 401--404, Beijing, China, 2000.

....previously developed in our group, to help detect and alleviate the presence of errors in speech recognition hypotheses. In the first method, an explicit out of vocabulary (OOV) word model is added into the model set of the recognizer in order to identify potential unknown words during recognition [1]. In the second method, the recognizer s hypotheses are post processed with a confidence scoring model in order to identify hypothesized words which may be misrecognized [5] Both methods attempt to identify the regions of an utterance where the recognizer cannot find reliable word hypotheses ....

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. of ICSLP, Beijing, 2000.

.... The speech recognizer is configured to use generic telephone based acoustic models, and is connected to the language understanding component via an # best interface [6] Since users may speak words which are not specified in the vocabulary, we have incorporated an out of vocabulary model [7]. Baseform pronunciations which do not occur in our large online dictionaries are generated by rule [8] SPEECHBUILDER provides an editing facility for developers to modify pronunciations. The recognizer deploys a hierarchical # gram grammar 2 Key Examples color red, green blue day Monday, ....

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. ICSLP, Beijing, 2000.

....acoustic models are used for processing the speech signal. These models are based on over 100 hours of training data collected primarily from the jupiter [37] voyager [13, 27] pegasus [30] and mercury [28] domains. The SpeechBuilder recognizer uses a new out of vocabulary (OOV) word model [2]. Out of vocabulary words, when not properly modeled, can seriously impair recognition. The reason for this is that an unknown word in a sentence is not only misrecognized, but can also result in deletion and substitution errors elsewhere in the utterance. Since SpeechBuilder recognizers are ....

I. Bazzi and J. Glass. Modeling Out-of-vocabulary Words for Robust Speech Recognition. In Proc. ICSLP, Beijing, 2000.

No context found.

I. Bazzi and J. Glass, "Modeling out-of-vocabulary words for robust speech recognition," Proc. of ICSLP, Beijing, 2000.

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