TELEPHONE-BASED CONVERSATIONAL SPEECH RECOGNITION IN THE JUPITER DOMAIN1 James R. Glass and Timothy J. Hazen Spoken Language Systems Group Laboratory for Computer Science Massachusetts Institute of Technology Cambridge, Massachusetts 02139, USA ABSTRACT This paper describes our experiences with developing a telephone-based speech recognizer as part of a conversational system in the weather information domain. This system has been used to collect spontaneous speech data which has proven to be extremely valuable for research in a number of different areas. After describing the corpus we have collected, we describe the development of the recognizer vocabulary, pronunciations, language and acoustic models for this system, and report on the current performance of the recognizer under several different conditions.
1. INTRODUCTION Over the past year and a half, we have developed a telephonebased, weather information system called JUPITER , which is available via a toll-free number for users to query a relational database of current weather conditions using natural, conversational speech2 . Using information obtained from several different internet sites, JUPITER can provide weather forecasts for approximately 500 cities around the world for three to five days, and can answer questions about a wide range of weather properties such temperature, wind speed, humidity, precipitation, sunrise etc., as well as weather advisory information. The JUPITER system makes use of our GALAXY conversational system architecture which incorporates speech recognition, language understanding, discourse and dialog modelling, and language generation . J UPITER has been particularly useful for our research on displayless interaction, information on demand, and robust spontaneous speech recognition and understanding. For example, since we attempt to fully understand all queries (i.e., not spot words), and do not constrain the user at any point in the dialog, it is crucial to have a high accuracy speech recognizer that covers, as much as possible, the full range of user queries. This paper describes our work in developing a robust recognizer in this domain. When the system was first deployed in late April 1997, the error rates of our recognizer initially more than tripled our laboratory baselines, due in part to the mismatch between the laboratory training and actual testing conditions. The real data had a much larger variation in environment and channel conditions (often with very poor signal conditions), as well as a much wider 1 This research was supported by DARPA under contract N66001-96-C-8526, monitored through Naval Command, Control and Ocean Surveillance Center. 2 In the U.S. and Canada: 888 573-8255 (http://www.sls.lcs.mit.edu/jupiter).
range of speakers (we had no children in our training data for example, and had mainly trained on native speakers without regional accents), speaking style (spontaneous speech vs. read speech), language (both for within-domain queries, and out-ofdomain queries), and other artifacts such as non-speech sounds and clipped speech due to the user interface (we do not currently allow for barge-in). As we have collected more data we have been able to better match the users’ vocabulary, and build more robust acoustic and language models. The result is that we have steadily reduced word and sentence error rates, to the point of cutting the initial error rates by over two thirds. In this paper, we describe the methods we have used to develop this recognizer, and report on the lessons we have learned in moving from a laboratory environment to dealing with real data collected from real users. Our experience has shown us clearly that while there is no data like more data, there is also no data like real data!
2.1. Data Collection Several different methods have been employed to gather data for the JUPITER weather information system. Beginning in February and March 1997, we created an initial corpus of approximately 3,500 read utterances collected from a variety of local telephone handsets and recording environments, augmented with over 1,000 utterances collected in a wizard environment . These data were used to create an initial version of a conversational system which users could call via a toll-free number and ask for weather information. The benefit of this setup is that it provides us with a continuous source of data from users interested in obtaining information. Currently, we average over 70 calls per day, and have recorded and orthographically transcribed over 59,000 utterances from over 10,560 callers, all without widely advertising the availability of the system. On average, each call contains 5.6 utterances, and each utterance has an average of 5.2 words. The data are continually orthographically transcribed (seeded with the system hypothesis), and marked for obvious non-speech sounds, spontaneous speech artifacts, and speaker type (male, female, child) .
2.2. Data Analysis Speaker Characteristics A breakdown of the live data shows that just over 70% of users are male speakers, with females comprising approximately 21% of the data, and children the remain-
der. A portion of the data was from non-native speakers, although the system performs adequately on speakers whose dialect or accent does not differ too much from general American English. Callers with strong accents constituted approximately 7% of the calls and 14% of the data. A very small fraction (0.1%) of the utterances included talkers speaking in a foreign language (e.g., Spanish, French, German, or Chinese). Signal Quality The signal quality of the data varied substantially depending on the handset, line conditions, and background noise. It is clear that speaker phones were used in approximately 5% of the calls due to the presence of multiple talkers in an utterance. Less than 0.5% of the data was estimated to be from cellular or car-phones. Non-Speech Sounds Over 11% of the data contained significant noises. About half of this noise was due to cross-talk from other speakers, while the other half was due to non-speech noises. The most common identifiable non-speech noise was caused by the user hanging up the phone at the end of a recording (e.g., after saying good bye). Other distinguishable sources of noise included (in descending order of occurrence) television, music, phone rings, touch tones, etc. Spontaneous Speech Effects There were a number of spontaneous speech effects present in the recorded data. Over 6% of the data included filled pauses (uh, um, etc.) which were explicitly modeled as words in the recognizer, since they had consistent pronunciations, and occurred in predictable places in utterances. Utterances contained partial words in another 6% of the data, although approximately two thirds of these were due to clipping at the beginning or end of an utterance. The remaining artifacts were contained in less than 2% of the data and included phenomenon such as (in descending order of occurrence) laughter, throat clearing, mumbling, shouting, coughing, breathing, sighing, sneezing, etc.
3. VOCABULARY The vocabulary used by the JUPITER system has evolved as periodic analyses are made of the growing corpus. The current vocabulary contains 1893 words including 638 cities and 166 countries. Table 1 shows a breakdown of the various types of words in the current vocabulary. Note that nearly half of the vocabulary contains geography related words. Type geography basic weather
Size 910 769 214
Examples Boston, Alberta, France, Africa I, what, January, tomorrow temperature, snow, sunny
Table 1: Categorical breakdown of the JUPITER vocabulary. The design of the geography vocabulary was based on the cities for which we were able to provide weather information, as well as commonly asked cities. Other words were incorporated based on frequency of usage and whether or not the word could be used in a query which the natural language component could understand. The 1893 words had an out-of-vocabulary (OOV) rate of 2.0% on a 2506 utterance test set. Table 2 shows some of the most frequently occurring OOV words along with some
Word N (letter) creek Los white mother news away stock
Count 10 9 7 7 6 6 5 5
Example Usage B A N G O R Bangor Oak Creek, Dayes Creek Los Los Angeles California white Christmas, white house I’m going to visit my mother what is the news of the day go away, get away from me I need a good stock tip
Table 2: Frequent OOV words in the JUPITER domain. can you do you excuse me give me going to you are
when is what about what are what will how about i would
never mind clear up heat wave pollen count warm up wind chill
Table 3: Examples of multi-word units in the JUPITER domain.
example usages of the word. Since the recognizer makes use of a bigram grammar in the forward Viterbi pass, several multi-word units were incorporated into the vocabulary to provide for greater long-distance constraint and, in some cases, to allow for specific pronunciation modelling. This would allow for explicit modelling of word sequences such as “going to” or “give me” to be pronounced as “gonna” or “gimme” respectively. Common contractions such as “what’s” were represented as multi-word units (e.g., “what is”) to reduce language model complexity, and because these words were often a source of transcription error anyway. Additional multi-word candidates were identified using a mutual information criterion which looked for word sequences which were likely to occur together. Table 3 shows examples of multi-word units found in the current JUPITER vocabulary.
4. PHONOLOGICAL MODELING In the current JUPITER recognizer, words are initially represented as sequences of phonetic units augmented with stress and syllabification information. The initial baseform pronunciations are drawn from the LDC PRONLEX dictionary. The baseforms are represented using 41 different phonetic units with three possible levels of stress for each vowel. The baseforms have also been automatically syllabified using a basic set of syllabification rules. After drawing the pronunciations for the JUPITER vocabulary from the PRONLEX dictionary, all baseform pronunciations were then verified by hand. Vocabulary words missing from the dictionary were hand coded. Alternate pronunciations are explicitly provided for some words. In addition to the standard pronunciations for single words provided by PRONLEX, the baseform file was also augmented with common multi-word sequences which are often reduced, such as “gonna”, “wanna”, etc. A series of phonological rules were applied to the phonetic baseforms to expand each word into a graph of alternate pronunciations. These rules account for many different phonological
phenomena such as place assimilation, gemination, epenthetic silence insertion, alveolar stop flapping, and schwa deletion. These phonological rules utilize stress, syllabification, and phonetic context information when proposing alternate pronunciations. We have made extensive modification to these rules, based on our examination of the JUPITER data. The final pronunciation network does not represent the words using the original 41 phonetic units utilized in PRONLEX. Instead, a set of 105 different units were used which include sub-phonetic, supra-phonetic and non-phonetic units in addition to standard phonetic units. For example, the recognizer treats most within-syllable vowel-semivowel sequences and some semivowel-vowel sequences as single units in order to better model the highly correlated dynamic characteristics of these sequences. Thus, the phonetic sequence [ow] followed by [r] is represented as a single segmental unit [or]. The recognizer also incorporates various non-phonetic units to account for nonlinguistic speech transitions and speech artifacts, silences, and non-speech noise. The 105 units also retain two levels of stress to augment each vowel unit. An example pronunciation graph for the word “reports” is shown in Figure 1. r
Figure 1: Pronunciation graph for the word “reports.” The arcs in the pronunciation graph can further be augmented with transition weights which give preference to more likely pronunciations and penalize less likely pronunciations. For JUPITER these weights were set using an error correcting algorithm on development data . This algorithm adjusted the arc weights in an iterative fashion in order to reduce the error rate of the recognizer on development data.
A class bigram language model was used in the forward Viterbi search, while a class trigram model was used in the backwards A∗ search to produce the 10-best outputs for the natural language component. A set of nearly 200 classes were used to improve the robustness of the bigram. The majority of the classes involved grouping cities by state or country (foreign), in order to encourage agreement between city and state. In cases where a city occurred in multiple states or countries, separate entries were added to the lexicon (e.g., Springfield, Illinois vs. Springfield, Massachusetts). Artificial sentences were created in order to provide complete coverage of all of the cities in the vocabulary. Other classes were created semi-automatically using a relative entropy metric to find words which shared similar conditional probability profiles. Table 4 shows examples of word classes. Since filled pauses (e.g., uh, um) occurred both frequently and predictably (e.g., start of sentence), they were incorporated explicitly into the vocabulary, and modelled by the bigram and trigram. Original orthographies were modified for training and testing purposes by removing non-speech and clipped word markers. When trained on a 26,000 utterance set, and tested on
raining snowing cold hot warm humid windy extended general
humidity temperature advisories warnings conditions forecast report
Table 4: Example word classes used in the JUPITER domain.
a 2506 utterance set the word-class bigram and trigram had perplexities of 18.4 and 17.1, respectively. These are slightly lower than the respective word bigram and trigram perplexities of 19.5 and 18.8. Note that the class bigram also improved the speed of the recognizer as it had 22% fewer connections to consider during the search.
The JUPITER system performs recognition using the segmentbased SUMMIT recognizer which can incorporate both segmentand landmark-based acoustic models . The nature of the acoustic models has varied over the course of system development, depending in large part on the amount of available training data. The current JUPITER configuration makes use of contextdependent landmark-based diphone models which require the training of both transition and internal diphone models. Internal diphones model the characteristics of landmarks occurring within the boundaries of a hypothesized phonetic segment, while transition diphones model the characteristics of landmarks occurring at the boundary of two hypothesized phonetic segments. Given the 105 phonetic units used in the JUPITER system, and the constraints of the full pronunciation graph, there were 4,822 possible diphone transition models and 105 internal models needed. We have explored two different methods of modelling transitions. The first method trained models for frequently occurring transitions, and used one “catch-all” model for remaining transitions. This method worked well, and was simple to train. We currently use a reduced set of 782 equivalence classes which were determined manually to insure that an adequate amount of training existed for each class and that the elements of each class exhibited contextual similarity. This method performs slightly better than the “catch-all” method. For each landmark, 14 MFCC averages were computed for 8 different regions surrounding the landmark, creating 112 different features. This initial feature set was then reduced from 112 features to 50 features using principal component analysis. The acoustic models for each class modeled the 50 dimensional feature vectors using diagonal Gaussian mixture models. Each mixture model consisted of a variable number of mixture components, dependent on the number of available training vectors for that class, with a maximum of 50 mixture components. The diphone models were trained on a subset of data which excludes utterances with out-of-vocabulary words, clipped speech, cross-talk, and various types of noise. The training data also excludes all speech from speakers deemed to have a strong foreign accent. The full set of within-domain training utterances used for acoustic modelling consisted of 20,064 utterances, which was 76% of the available data at the time.
7. EXPERIMENTS Over the course of the past year the JUPITER recognizer has had a steady improvement in its performance; this has been a result of both an increase in training data and improvements to the system’s modeling techniques. The test data consists of sets of calls randomly selected over our data collection period. The current test set consists of 2506 utterances, of which 1806 were considered to be “in domain” as they were covered by the vocabulary, were free of partial words, crosstalk, etc. Of these sentences, 1290 were from male speakers, 274 from females, and 242 from children. Table 5 shows the performance of the JUPITER recognizer on this test set using word error rate (WER) and sentence error rate (SER) as the evaluation metrics. As can be seen in the table, the system tends to perform reasonably when it encounters queries spoken by adults without a strong accent, that are covered by the domain, and that do not contain spontaneous, or non-speech artifacts. Females had 50% more word errors than males, while children had 300% more word errors than males. This is probably a reflection of the lack of training material for females and children. The system has considerable trouble (64.5% WER) with “out of domain” utterances containing out-of-vocabulary words, partial words, crosstalk, or other disrupting effects. This rate is artificially high, however, due to the nature of the alignment procedure with reference orthographies (e.g., partial words always cause an error for example, due to the nature of our mark-up scheme). Table 5 also shows the performance on speakers judged to have strong foreign accents, who were not included in the standard test set. These data consisted of 3,225 in-domain utterances, and had an error rate more than double the baseline in-domain error rate. Finally, we also evaluated the recognizer on “expert” users (i.e., mainly staff in our group) who have considerable experience using the JUPITER system, but were not used for training. The system had extremely small error rates for these users. This behavior is typical of users who become familiar with the system capabilities (a case of users adapting to the computer!). Test Set Standard In domain Male (In domain) Female (In domain) Child (In domain) Out of domain Non-native (In domain) Expert (In domain)
# Utts 2506 1806 1290 274 242 700 3225 354
WER 24.0 11.8 8.4 12.3 25.0 64.5 26.6 2.1
SER 41.9 26.6 22.2 28.5 48.3 81.4 56.7 7.9
Table 5: Current JUPITER performance on various test sets.
8. DISCUSSION & FUTURE WORK The speech recognizer described in this paper is only one component of the full JUPITER conversational system [8, 11]. The current interface between the recognizer and our language understanding component is via an N -best interface. Although we have reported only first-choice error rates in this paper, the understanding error rates are typically better, since many word confusions do not impact understanding.
There remain a considerable number of ongoing areas of research we are presently pursuing, which should help improve performance. Recent developments in probabilistic segmentation , near-miss modelling , heterogeneous classifiers , and tighter integration of linguistic knowledge , have shown improvements in our JUPITER baseline, although they have not yet been propagated to the data collection system. We have also been able to significantly speed up the system (which is constrained to near real-time performance) using a more flexible finite-state transducer representation under active development. The system to date has used a pooled speaker model for all acoustic modelling. It should be possible to achieve gains through speaker normalization, short-term speaker adaptation, and better adaptation to the channel conditions of individual phone calls. Adaptation may also be useful to help improve performance on non-native speakers. Since a phone call might have multiple speakers, we are exploring within-utterance consistency techniques which have given us gains elsewhere . The data collection efforts have produced a gold-mine of spontaneous speech effects which are often a source of both recognition and understanding errors. For example, partial words typically cause problems for the speech recognizer. Another source of recognition errors is out-of-vocabulary words, which are often cities not covered in the vocabulary. These issues have caused us to begin work in confidence scoring, which was an area we had not previously addressed . Finally, we plan to explore the use of dynamic vocabulary and language models, which may help to alleviate some of the unknown city problems. Acknowledgments The JUPITER data collection effort was coordinated by Joe Polifroni, and heroically transcribed by Sally Lee. Jon Yi syllabified the LDC PRONLEX dictionary.
9. REFERENCES 1. J. Chang, Near-Miss Modeling: A Segment-Based Approach to Speech Recognition. Ph.D. thesis, MIT, May 1998. 2. G. Chung and S. Seneff, “Improvements in speech understanding accuracy through the integration of hierarchical linguistic, prosodic, and phonological constraints in the JUPITER domain,” in these Proceedings, Sydney, 1998. 3. J. Glass, J. Chang, and M. McCandless, “A probabilistic framework for feature-based speech recognition,” in Proc. ICSLP, Philadelphia, PA, pp. 2277–2280, 1996. 4. A. Halberstadt and J. Glass, “Heterogeneous measurements and multiple classifiers for speech recognition,” in these Proceedings, Sydney, 1998. 5. T. J. Hazen, The Use of Speaker Correlation Information for Automatic Speech Recognition. Ph.D. thesis, MIT, January 1998. 6. S. Lee and J. Glass, “Real-time probabilistic segmentation for segmentbased speech recognition,” in these Proceedings, Sydney, Australia, 1998. 7. C. Pao, P. Schmid, and J. Glass, “Confidence scoring for speech understanding,” in these Proceedings, Sydney, Australia, 1998. 8. J. Polifroni, S. Seneff, J. Glass, and C. Pao, “Evaluation methodology for a telephone-based conversational system,” in Proc. First Int’l. Conference on Language Resources and Evaluation, Granada, Spain, pp. 43–49, 1998. 9. S. Seneff, E. Hurley, R. Lau, C. Pao, P. Schmid, and V. Zue, “Galaxy-II: A reference architecture for conversational system development,” in these Proceedings, Sydney, Australia, 1998. 10. V. Zue, J. Glass, D. Goodine, M. Phillips, and S. Seneff, “The SUMMIT speech recognition system: Phonological modelling and lexical access,” in Proc. ICASSP, Albuquerque, NM, pp. 49–52, 1990. 11. V. Zue, S. Seneff, J. Glass, L. Hetherington, E. Hurley, H. Meng, C. Pao, J. Polifroni, R. Schloming, and P. Schmid, “From interface to content: Translingual access and delivery of on-line information,” in Proc. Eurospeech, Rhodes, Greece, pp. 2047–2050, 1997.
Telephone-Based Conversational Speech Recognition in the ...
TELEPHONE-BASED CONVERSATIONAL SPEECH RECOGNITION IN THE JUPITER DOMAIN1 James R. Glass and Timothy J. Hazen Spoken Language Systems Group Laboratory ...
Sphinx-4 is one of the ASR systems whose code is freely available for download . II. THE CMU SPHINX. The Sphinx system has been developed at. Carnegie ...
Medical One (DMO), is available for direct use by Windows users and through virtualization software for. Mac and mobile .... receive a âWelcome Letterâ which includes installation and log on instructions. The letter ..... [Hint for Epic users: Ep
software is installed (a 20 minute process), you must create a new âuser profileâ by teaching dragon how you talk. You are required to ... out of the procedure documentation and insert it any time you desire (with any patient) by simply speaking
Sep 29, 2017 - Instead, we spent the vast majority of our time not even on drills, but on game related scenarios. We played a lot of defence and a lot of open field, dynamic attack scenarios. Sometimes we'd play full contact, full-field, but often we
3.2 Test results with Sphinx4â¦ ... Chart 3.3 Experiment Results with Sphinx-4 Audio Inputâ¦ ..... Download binary distribution of Sphinx4 from the below link.
Based on Normalized Relevance Distance ... advantage of Normalized Relevance Distance [Schaefer et .... working with Normalized Information Distance.
features and DTW to compare the Tamil words and numerals test patterns. .... To determine the "meaning" of this voice input, the computer attempts to ..... Cummins,âSpeaker Identification Using Instantaneous Frequenciesâ ,IEEE Transactions.
of having a cold on the German version of the Wisconsin. Upper Respiratory Symptom Survey (WURSS-24) . The questionnaire is an evaluative ...
Data collection and Gaussian classification (20%) You will collect some speech data (from yourself) and perform a small study on vowel classification using just one feature vector per ..... You should also look inside your decision algorithms and see
without compromising the recognition accuracy on in-vocabulary words. .... Among others, de Marcken , Brent , Goldsmith , Creutz and La-.
Although we will try to go beyond raw data by outlining our motivation, and drawing conclusions that reach further, our main focus will be on accuracy scores and error rates. The main question we ask will be whether the proposed methods improve the r
the estimated mixture weight for the ith GMM component. In our system, since PMC technique cannot be applied to the. DMFCC feature, noise-dependent HMMs ...
features and DTW to compare the Tamil words and numerals test patterns. .... To determine the "meaning" of this voice input, the computer attempts to ..... Cummins,âSpeaker Identification Using Instantaneous Frequenciesâ ,IEEE Transactions.
The Introduction of the speech should do three things: 1. Get the ... The Body of the speech can be well planned with: 1. ... begin writing the body of their speech.
Oct 16, 2012 - free speech doctrine to the phenomenon of campus hate speech in order to draw out a cause of action .... 18 See Craddock, supra note 17, at 1048; Lawrence Friedman, Regulating Hate. Speech at Public ...... L. Wkly., Feb. 18,. 2011, at
by the Governor to turn the prisoners over to the authorities of Idaho. ...... in this if he had not sought to save his life before he turned to saxe his soul. But there are certain ...... tangle him up I nrould have been rery glad if they had put St
Rachelle Brooks. Guilbert Brown. Jennifer Brown. Keith Brown. Norman Bryan. Jennifer Buckley. Danielle Buehrer. Katie Busby. Scott Busche. Christina Butler ... William Fendley. Salma Ferdous. James Fergerson. Ann Ferren. Norma Fewell. Bamby Fields. J
2013) as well as the consistent opposition of the nationalist party Milliyetçi ...... ://www.systemdynamics.org/conferences/2005/proceed/papers/AKCAM225.pdf.
Aug 13, 2007 - verry, wee, younger, yw  alas, ale, anyhow, anythin, appetizers, asl, awhile, balancing, bil, boat, bottles, brown, chease, chessecake, currently, ... recommendation, review, sa, scare, seeya, sexual, shove, shrimp, sil, similar, s