SRILM is a collection of C++ libraries, executable programs, and helper scripts designed to allow both production of and experimentation with statistical language models for speech recognition and other applications. SRILM is freely available for noncommercial purposes. The toolkit supports creation and evaluation of a variety of language model types based on N-gram statistics, as well as several related tasks, such as statistical tagging and manipulation of N-best lists and word lattices. This paper summarizes the functionality of the toolkit and discusses its design and implementation, highlighting ease of rapid prototyping, reusability, and combinability of tools. 1.

In large-vocabulary, multi-pass speech recognition systems, it is desirable to generate word lattices incorporating a large number of hypotheses while keeping the lattice sizes small. We describe two new techniques for reducing word lattice sizes without eliminating hypotheses. The first technique is an algorithm to reduce the size of non-deterministic bigram word lattices. The algorithm iteratively combines lattice nodes and transitions if local properties show that this does not change the set of allowed hypotheses. On bigram word lattices generated from Hub4 Broadcast News speech, it reduces lattice sizes by half on average. It was also found to produce smaller lattices than the standard finite state automaton determinization and minimization algorithms. The second technique is an improved algorithm for expanding lattices with trigram language models. Instead of giving all nodes a unique trigram context, this algorithm only creates unique contexts for trigrams that are explicitly represented in the model. Backed-off trigram probabilities are encoded without node duplication by factoring the probabilities into bigram probabilities and backoff weights. Experiments on Broadcast News show that this method reduces trigram lattice sizes by a factor of 6, and reduces expansion time by more than a factor of 10. Compared to conventionally expanded lattices, recognition with the compactly expanded lattices was also found to be 40 % faster, without affecting recognition accuracy. 1 1.

Our research concerns spoken language understanding within the domain of automated telecommunication services. In the recent papers we presented a new methodology for training of statistical language models for recognition and understanding of utterances from large corpora of phone sequences obtained as the output of a task-independent ASR-system. The advantage of this strategy compared to the traditional word-based strategy is that we don't have to manually transcribe large amounts of data in order to extract acoustic morphemes to train the classifier. Since the baseline strategy suffered high False Rejection Rates caused by finding no acoustic morphemes in the test data, we describe in this paper how approximate matching can be incorporated in the Bayes-classifier to reduce FRR. The experiments are evaluated for "How May I Help You?"-task.

Word Lattices are used by most speech recognizers as a compact representation of a set of alternative hypotheses. In large-vocabulary, multi-pass recognition systems it is important to generate word lattices incorporating a large number of hypotheses but at the same time keeping the size of the representation as small as possible. Previously we presented a method for identifying mutually supporting and competing word hypotheses in a recognition lattice. In this paper we show how the outcome of this method can be used for compressing lattices. The success of the new technique comes from the ability to discard links with low a posteriori probability and recombine the remaining ones to create a new set of hypotheses. Experiments on the Switchboard corpus show that this method results in better compression results than the conventionally used technique.

"Go forth and make masterpieces of hyphenation patterns... "

Larr vocabular y continuous speech reech ition can benefitfre an e#cient data strR turfor rrR/sentingalarE number of acoustic hypotheses compactly. Wor gr1:1 or lattices have been chosen as such an e#cientinter face between acousticroust ition engines and subsequent languageprguag ing modules. This paper firR investigates the e#ect ofprEI/-- dur ing acoustic decoding on the quality ofwor lattices and shows that by combiningdi#erEE pre ing options (at the model level and wor level), we can obtain wor lattices withcompar bleaccurE/ to theorRE/ al lattices and a manageable size. In orer to use the wor lattices as the inputfor a post-prt-RI ing language module, they shouldprx--:/1 thetar/E hypotheses andtheir scor while being as small as possible. In this paper weintr oduce awor grC comprmpR/-- algor thm that significantlyrnt ces the number ofwor-- in thegrRxEE alrRx---- entation without eliminatingutter ance hypothesesor distortRI their acousticscort . Wecompar this wor grR comprCx/)R algor thm withsever lother latticesize-rRI cing appr aches and demon strnR thereRx1C-- strx gth of the new wor gr1/ comprw sionalgor:I+ for decr: ing the number ofworC in thereR/) entation. ExperR entsar conductedacrRI corRI/ and vocabular sizes todeterE/R the consistency of theprR/--) and comprC sionrnRIIC) # 2003 Elsevier Science Ltd. AllrlRI srEIE ved. 1.I5k4 Wor latticesar often chosen as theinter/C1 between an acousticrusticRx-- and a subsequent prubsequ using amor complex language model (LM)or mor specific acoustic model because of www.elsevierw.elsevi te/csl COMPUTER SPEECH AND LANGUAGE * Corr)R)R)Rr author Tel.: +1-765-494-3652; fax: +1-765-494-3371. E-mailaddr9(--)b harRxC/1:Rwxxx/Rrx+ yangl@ecn.purxxx/Rr (M.P.Har.RIC mike.johnson@marrx+Rwxx (M.T. Johnson),lhj@ecn.pur)xRwEE...