You are on page 1of 9

RC23301 (W0408-092) August 19, 2004

Computer Science

IBM Research Report
Safety Driver Manager

Dimitri Kanevsky, Barbara Churchill, Alex Faisman,
David Nahamoo, Roberto Sicconi
IBM Research Division
Thomas J. Watson Research Center
P.O. Box 218
Yorktown Heights, NY 10598

Research Division
Almaden - Austin - Beijing - Haifa - India - T. J. Watson - Tokyo - Zurich

LIMITED DISTRIBUTION NOTICE: This report has been submitted for publication outside of IBM and will probably be copyrighted if accepted for publication. Ithas been issued as a Research
Report for early dissemination of its contents. In view of the transfer of copyright to the outside publisher, its distribution outside of IBM prior to publication should be limited to peer communications and specific
requests. After outside publication, requests should be filled only by reprints or legally obtained copies of the article (e.g., payment of royalties). Copies may be requested from IBM T. J. Watson Research Center , P.
O. Box 218, Yorktown Heights, NY 10598 USA (email: reports@us.ibm.com). Some reports are available on the internet at http://domino.watson.ibm.com/library/CyberDig.nsf/home.

possible variants for speaking commands are Studies of road safety found that human error was the kept in special grammar files (one file for each topic or sole cause in more than half of all accidents (see for application). In our paper we analyze some aspects of this IBM Research. All driver activities in cars (driving. If it finds an appropriate command in some file. issuing commands via introduced novel solution for using a speech interface to voice).) contribute to driver workload. NY 10598 kanevsky@us. provide for their cars. It has been commands in a natural manner rather than having to shown in various experiments that well designed voice remember one or two variants supported by the system. and are not telephone from ringing or interrupt a dialog between the robust. internet access) are becoming increasingly devoted to a discussion of these and other aspects of popular. information processing ([7]).g. such as "What is the distance to JFK?" or "Or talking over telephone. When the system gets a voice response. this workload manager may prevent the available everywhere. syntax. Some car manufacturers see advantages in using risk factors in its safety assessment of the environment . The workload manager would then adjust chips). In our 1.. NLU components mail message or has had a high number of traffic should be located on a server that is accessed by cars accidents in the past.g. but they may distract drivers from their main driver safety. as a mechanism to increase driver One of the ways to address driver safety concerns is to safety. cellular contribute to additional driving errors. It extends concepts described in the paper [3]. driving tasks and negatively affect driving safety. Voice control becomes more efficient in develop an efficient system that relies on voice instead reducing driver distraction if drivers can speak of hands to control Telematics devices. J. Safety Driver Manager Dimitri Kanevsky . One of the ways to reduce a powerful due to cost considerations. Alex Faisman. recalling specific falling asleep. This paper is telephone. if a driver in a nearby car is listening to an e. control interfaces can reduce a driver’s distraction In this paper we discuss some ways to increase compared with manual control situations (see for naturalness. [16]. To address this. this “heightened risk” information remotely or NLU should be downsized to run on local could be sent anonymously to the workload manager in computer devices (that are typically based on embedded another car. This paper addresses some aspects of voice user 2. upgraded NLU and speech processing on the client in In response to the heightened risk caused by the the car. One of the reasons why humans commit searches through files (starting with the most relevant so many errors lies in the inherent limitation of human topic). With the increase in it executes the command. nearby cars who are at component .com more information that drivers need to process and more Abstract devices that drivers need to control that might Telematics services in cars (like navigation. Our department is developing a “quasi-NLU” driver and a car system in other. CIT speech systems can significantly problem. Another aspect of this paper is a recently when playing voice games. Therefore difficult problem that typically requires significant sharing some information about other cars and driver computer resources that are usually not available in conditions could lead to increased driving safety.a “reduced” variant of NLU that can be run higher risk because of the nearby “offending” car. it example.). Roberto Sicconi IBM T. since remote connections to servers are not offending car. Otherwise the system popularity of Telematics services in cars (like executes other options that are defined by a Dialog navigation. Yorktown Heights. Barbara Churchill. Introduction approach. we introduce the idea of a distributed improve a driver-vehicle relationship and contribute to user interface between cars. internet access) there is Manager (DM) . Designing This fact led to the development of Conversational workload management in a user interface is a difficult Interactivity for Telematics (CIT) speech systems at task. in CPU systems with relatively limited resources. It is difficult for a driver to remember a reduce driver drowsiness and prevent a driver from complex speech command menu (e. Watson Research Center. It is well known that the driving safety.ibm. Computers in cars are usually not very example [14] or [5]). cellular telephone. David Nahamoo. etc. For local computer processors that car manufacturers example. The DM component is a rule based 1 . Finally. controlling Telematics devices. solutions. But the development of full fledged safety of a driver depends not only on the driver himself Natural Language Understanding (NLU) for CIT is a but on the behavior of other drivers nearby. how far is JFK?" or "How long to drive to JFK?" etc. Low CPU driver’s cognitive workload is to allow the driver to resources are a limiting factor for embedded speech speak naturally when interacting with a car system (e. can have delays. Voice control interface interface in cars.

language models) related to similarities interact with a CIT system. 3. CIT Dashboard and other data and provides means to In this system the front-end computes standard 13- calculate cognitive weights. accurate. in New York improved and adjusted to improve Driver Safety 2. (DM) Partly cloudy. developed in the ViaScribe speech interactive there is significant need for a low-resource speech framework ([2]). It is difficult to collect sufficient data that Safety Driver Manager components (e. Fig. that allows users to easily add new phrases and commands if they find that the system Car computers are usually not very powerful due to cost does not understand the phrases they are using. No matter how large the between users may also be carried out in any of a wide data collection is. NY. what did you say? statistical language models can be created from these . (Driver) Please. A high view of the architecture of the recognition system that is robust. the labeler and the decoder. some users will produce some phrases range of suitable methods. Examples of LT tasks are as follows: reduce task complexity for the NLU system. 2. When processing speech. (Driver) When will we get there? sufficient number of phrases has accumulated. Ask questions (via a text to speech module) to relevant data in databases. then . (Driver) What is the weather forecast for In particular. Currently this issue is addressed by a convenient user interface that is provided by a research prototype of a CIT 3. 10% chance of rain drivers’ dialogs or from more sophisticated automatic (The DM assumes that the driver means Yorktown. In particular. resolve ambiguities: • Generate and learn patterns from stored data. LT could .) and a driver to system (LT). including those described in that are not represented in the collected data nor in [11] and [12].g. etc. to which other parts of the framework can be connected. grammars that are developed from this data. 1 considerations. (DM) Partly sunny. Text To Speech (TTS). ambiguous references from context: . advanced DM should be able to perform: • Extract and record the Driver Safety Manager 1. in which Speech An example of a low-resource system that is executed Engines (SE). The growing necessity of the shows a snapshot of an interface that allows users to conversational interface demands significant advances write new phrases and commands. such as how well phrases in files cover multiple ways to speak disclosed in [10] . The communication is provided by an Engine Logically a speech system is divided into three primary Manager component that is designed as a message bus. the adaptation of commands. modules: the front-end. Various a brief description of the IBM embedded speech Dialog Managers (DM) have been developed using recognition system that is based on the paper [4]. the problem dimensional mel-frequency cepstral coefficients 2 . NLU. language models and NLU processors. When a . speech fully represents all possible ways that users might recognition. augmented by 1 MB or less capturing/recording engines and some of the other of DRAM can be found in [2].) grammars for domains related to such created topics. when the phrases in some file become very large (which might driver asks the inexplicit question “What about lead to increased speech recognition error rates). Examples of some technology that can be used for such failure. (Driver) No. Manage failure and provide contextual. and efficient. For example. plot a course to Yorktown • Learn from this data how Safety Driver . the computational workload is A workload manager (WM) receives sensor data from divided approximately equally among these modules. Java. such as files which include typical phrases . LT should be able to modify NLU today? components. Embedded speech recognition framework. A wide range of known mechanisms may be employed The quality of this “quasi-NLU” module depends on to promote interactions of LT with drivers. (DM) Sorry. telephone directories. The • Monitor driver and passenger actions in the following are examples of conversations between a car’s internal and external environments across a driver and DM that illustrate some of tasks that an network. Fill in missing information and remove Manager performance and improve driving safety. Also.dependent help and actions topic identification are provided in [9]. In the future. . On the other hand. (Driver) What about Ossining? add new phrases to NLU files that it finds from some . e-mail systematically by creating a learning transformation systems.sub-system that can interact with the car and external of instantaneous data collection could be dealt systems (such as weather forecast services. then Ossining?” it assumes that the driver is still asking LT could split files by topics and adapt or create new about weather. The CIT interface is in processing power on the one hand. audio by a 50 DMIPS processor. (DM) Within Massachusetts? Manager components and driver behavior could be . 50% chance of rain that are linked with commands. (Driver) I asked when will we get there. If the number of from the earlier conversational context. corpora and augmented with grammar-based processors. and speech and a special case of a general interactivity system that is natural language technologies on the other. In what follows we give components have been developed using C/C++. CIT framework is shown in Fig.

7. such as car minimize the likelihood of the training data given all noise. 3 . The MMI estimation process that above 200 Hz. A standard 10 ms frame-shift was chosen since it reduces discriminative training procedure was applied to the overall computational load significantly without estimate the parameters of these phones. Each frame of attempts to simultaneously (i) maximize the likelihood speech is windowed with a Hamming window and of the training data given the sequence of models represented by a 13 dimensional MFCC vector. can offer. We corresponding to the correct transcription. drivers should be presented with yields a ranked list of the top 100 HMM states. and the the states on the basis of their likelihoods. media. The acoustic model drowsiness by detecting drowsiness via speech comprises context-dependent sub-phone classes biometrics and. and may contribute to safety. with diagonal higher level of interaction with a driver than current covariance matrices to give an initial acoustic model. external stimuli that are interactive to improve their Likelihoods are inferred based upon the rank of each alertness. and then ranking awareness-state of the driver will be monitored. it received favorable worldwide media its right. system to noise. by increasing arousal via (allophones). the invention [8] suggests fighting Markov model (HMM) approach. will be presented with vocabulary. The 24 triangular mel-filters are was used in this work is described in [6] and [15]. that increase driver procedure are run to obtain the final baseline model.1% for 60 mph). The purpose of Artificial vectors that collect at each leaf are modeled by a Passenger part of the CIT project at IBM is to provide a Gaussian Mixture Model (GMM).g. 2001. Currently. MMI training affecting the recognition accuracy. phonemes. and interactive books.000 vehicular accidents vectors. The front-end also performs adaptive mean removal and adaptive energy normalization to 4. is obtained by computing the likelihood of the acoustic In the Artificial Passenger (ArtPas) paradigm. Words more stimulating content than drivers who appear to be are represented as sequences of context-dependent alert. newspaper and radio that vectors and specifying the terminal nodes of the tree as Artificial Passenger was perceived as having the the relevant instances of these classes. Discarding the low testing conditions (e. It is a common experience for self loop and a forward transition.(MFCC) from 16-bit PCM sampled at 11. The labeler computes first and second differences of the 13-dimensional cepstral Fatigue causes more than 240. The options. the output distributions on the state focus. Driving. The allophones are identified by growing a attention. the vector using the GMM at each state. The rank of an HMM state reading jokes. hidden direction. The context for a given phone is speech interactivity. Among the efforts presented in this general system uses the familiar phonetically-based. drivers who are alone in a elements to yield a 39-dimensional feature vector. speech evaluation experiments yields Hz]. with center frequencies equally spaced in the improvement from 20% to 40% relatively depending on corresponding mel-frequency scale. and (ii) empirically observed that noise sources. have significant energy in the low frequencies possible sequences of models allowed by the grammar and speech energy is mainly concentrated in frequencies describing the task . Drivers implements a synchronous Viterbi search over its active evidencing fatigue. Ideally. When the patent was granted in composed of only one phone to its left and one phone to May 22. asking questions. Each allophone is potential to dramatically increase the safety of drivers modeled by a single-state Hidden Markov Model with a who are highly fatigued. Often these vector according to observation densities associated do not provide sufficient stimulation to assure with the states of the system's HMMs.5500 2001 . This can include both conversational dialog and transitions are expressed in terms of the rank of the interactive games. HMM state are conditioned upon one phone of left shift of 15 ms. Driver drowsiness prevention – reduce the effects of channel and high variability in the Artificial Passenger signal levels respectively. for example. with each phoneme modeled as a three-state Front-End Processing Speech samples are partitioned HMM. The sequence of and hands. such as CD players or radio stations. The observation densities associated with each into overlapping frames of 25 ms duration with a frame. This computation wakefulness. thereby limiting most current interactive rank likelihoods is then forwarded to the decoder. It became clear from the numerous press decision tree using the context-tagged training feature articles and interviews on TV.025 KHz. This could enhance the driver experience. In awareness and help to determine if the driver is losing our system. which may be changed dynamically. Starting with these initial sets of GMMs several This is envisioned as a series of interactive modules iterations of the standard Baum-Welch EM training within Artificial Passenger. however. and concatenates these with the original every year. using voice only. occupies the driver’s eyes HMM state by a table lookup ([1]). The training feature drivers to talk to other people while they are driving to vectors are poured down the decision tree and the keep themselves awake. the GMM modeling the leaf. In therefore placed in the frequency range [200Hz . A 15 ms frame-shift instead of the context and one phone of right context only. The vehicle have access only to media such as music and labeler then computes the log likelihood of each feature radio news which they listen to passively. The scenarios for HMM state instead of in terms of the feature vector and Artificial Passenger currently include: quiz games. The decoder content will be modified accordingly.6% error rate for 0 speed and frequencies in this way improves the robustness of the 10. if needed.

radio. Maximum Joe (startled): It’s not her birthday! load would be represented by the position where there is ArtPas: You seem a little tired. The ArtPas: You were a wiz at “Name that Tune” workload manager is closely related to the event last time. Joe: Hall of Fame. The relationship between a driver and a car typical scenario involving Artificial Passenger: should be consistent with the information from sensors. The workload manager should be designed in such a way that it can integrate sensor information and rules on Imagine. and arousal levels. like Joe: 500 computing how often the driver answered correctly and ArtPas: I see. cortical activation.) levels are higher (reacts to galvanic skin responses such and actions that a driver can perform in cars via voice as temperature and humidity). laughing create higher arousal levels than sitting quietly. orthogonally distinct axis. beyond which there would be “overload”. Sleep is receding to additional stimuli and on what types of stimuli should the edges of his mind. etc. based on his/her answers By the time Joe has answered a few questions and has as well as on information from the workload manager. If Joe loses his concentration on be provided (e. heavy rain).The Artificial Passenger interaction is founded on the 5. It interprets the status of a driver’s alertness. verbal stimuli via speech applications the game (e. does not respond to the questions which or physical stimuli such as a bright light. Do about traffic conditions on an expected driver route). the use of sensory technology in cars.g. low) and In this section we provide a brief analysis of the design valence (positive. loud noise. These sensors information retention and memory.) and whether to suggest to a driver to stop for rest. There is a conditions. human voices and faces (especially if larger application of brakes. speaking speed. Want to manager that detects when to trigger actions and/or try your hand at trivia? make decisions about potential actions. and lot of research into what factors increase psychological prioritizing message delivery. His head starts to nod … surface”. Want to play a game? both maximum vehicle stress and maximum driver Joe: Yes.g. movement. changing lanes). Driver Manager. eyelid status). Arousal reflects the level of driver Safety Manager (see Fig. The Artificial The system permits the use and testing of different Passenger can detect that a driver does not respond statistical models for interpreting driver answers and because his concentration is on the road and will not information about driver conditions.g. snakes) increase information about local events (e. It been engaged with the dynamics of the game. The system uses Joe: OK. music. physical stimulus (e. Most well known emotion researchers agree that arousal (high. provide than life). It respiration. you want the 100. a set of rules for starting and stopping the interactions ArtPas: Pick a category: Hollywood Stars. what did you get your would represent stress on the driver. 4 . Workload Manager concept of psychological arousal. his will make decisions on whether the driver needs activation level has gone way up. An object of the stimulation of the person as measured by physiological workload manager is to determine a moment-to-moment aspects such as heart rate. verbal alarm). negative) are the two fundamental of the workload management that is a key component of dimensions of emotion.g. ArtPas: Hey Joe. Artificial Passenger asks) the system will activate a etc. Hedging your bets are you? the length of delays in answers. and analysis of the user's cognitive workload. accomplishes this by collecting data about user they have to have a very low level of arousal. stress. One axis of the surface would represent stress on the vehicle and another. Values on each daughter for her birthday? axis could conceivably run from zero to one. the driver’s cognitive workload. We also know that speaking and provide information about driver characteristics (e. I was impressed. driver “Joe” returning home after an extended when and if distracting information is delivered. 3). which glows when arousal to the driver (e.g. The following is a control.g. It controls interruption of a dialog Magic Moments or Hall of Fame? between the driver and the car dashboard (for example. interrupting a conversation to deliver an urgent message ArtPas: I bet you are really good at this. e-mail etc. On longer trips the workload manager is connected to a driving risk Artificial Passenger can also tie into a car navigation evaluator that is an important component of the Safety system and direct the driver to a local motel or hotel. (or interventions). The goal of the Safety Driver Manager is to evaluate the potential risk of a traffic accident by producing measurements related to stresses on the driver and/or vehicle. phone rings. We know that allow for the monitoring of driver actions (e. monitoring local and remote events. and scary images (fires. 500 or 1000 dollar It can use answers from the driver and/or data from the level? workload manager relating to driver conditions. There is also growing Arousal levels can be measured fairly easily with a amount of distracting information that may be presented biometric glove (from MIT).g. For someone to be sleepy or fall asleep. There is rapid growth in arousal since this can result in higher levels of attention. The driver distract the driver with questions. This business trip during which he had spent many late can be designed as a “workload representational nights.

unique needs. For example. etc. facilitated by a particular driver’s repeated routines. because they are near or are going to pick up their kids in time by. Privacy Manager. The social driver is playing a voice interactive game. it is required that Addressing privacy concerns: The safety driver drivers younger than 17 have a mark provided on their manager framework should be designed such that it will cars to indicate this.g. sense of humor. in the Artificial Passenger music to play. biometrics) and as output it responses cannot be intercepted. be straightforward for the application designers to it was required that deaf or hard of hearing drivers 5 . In Russia (at least in Soviet times). if a driver does not laugh at certain type Driver speaks over the phone AND the car moves with of jokes. for reflected in Fig. The important input to the protect the end user’s privacy. it is example of such an option is the opportunity for drivers hot in a car. a baby in a car.g. Abstract situations system is. and what could include. another car approaches on the left AND the scores from the same town. An car windows are closed. some drivers might behavioral patterns. Distributive user interface between cars driver rarely visits other states. whose task is to recognize critical situations. Insurance) are can be resolved through situation recognition. find it disconcerting if the Artificial Passenger system and prioritize message delivery. but they do not “learning” habitual patterns and goals. the games. The social impact of the system will be managed the midst of locking a car.. etc. a baby is in a car is being developed to enable service providers to AND there are no people in a car. Some elements of this (geographical) state or municipality. such that the driver’s network data. another car is enhance the end user’s driving experience. attenuate the risks presented to other drivers who may the system could assist in determining whether drivers be in “high risk” situations. driver asks for a feedback in one area (football scores from a driver’s stock quotation AND presses brakes AND it is raining home town) leads to additional related content (baseball outside. For example. illness. 3. then that type will not be presented. passing a car where the driver is distracted by playing reordering a path from the cleaners. a car is located to be informed about the existence of the Artificial in New York City. and the used by different modules such as workload. a key is left on a car seat. listening to books or engaging in a telephone grocery store. Situation information can be driver’s conversational style. In some countries. to they need to do to maximize its performance to suit their cleaners. the uncertainty resides in the lack of an expressed Telematics can be found in [13]. weather. e. The uncertainty approach (e.environmental factors. the mall. privacy and/or the types of content presented. Privacy and social aspects in other cars. a driver is on Passenger capability. concerns associated with Artificial Passenger can be The dialog manager may well at times require addressed by allowing the users to specify their desires uncertainty resolution involving complex situations. a key is left on a car seat AND a driver is in effect. car sensor data. The system will include a analysis of the driver's cognitive workload. for example: Goals: get to work. what the system can and cannot do. conversation. Existing technologies can attenuate the risks to a driver which provides a good opportunity for the system’s in managing his/her own vehicle. The concept associated The safety of a driver depends not only on the driver with learning driver behavioral patterns can be himself but on the behavior of other drivers near him. when the opens the window for them automatically. Situations could be simple. to a movie.). by systems that learns driver Artificial Passenger system (e. others might workload manager performs a moment-to-moment find this a key feature. with clear instructions about how the way home. the driver is diabetic AND by making sure that users clearly understand what the has not taken a medicine for 4 hours. So. Positive high speed AND the car changes lanes. there are no people in a car. It receives through a carrier's network. example. This can be achieved produces a list of situations. a driver has diabetes. and the approaching AND the driver is looking in the opposite system will be designed to ensure that it has this desired direction.g. It would thus appear helpful at times to inform a driver about such risks associated with drivers 6. video. Driver history: paradigm the interaction can be customized to suit the accidents. This should include workload manager is provided by the situation manager encryption of the message traffic from the vehicle. for to turn it off if they opt not to use it. perhaps.). visits. for instance. provide driver distraction detection. restaurants. the driver’s eyes are closed. social and legal issues in Here. Server or Portal Server. shops. allowing the end user an for instance: a dog locked in a car. as and requirements through the subscriber management exemplified in the following verbal query by a driver: tools. provider’s secure environment. Simple situations could include. etc. example: a dog locked is locked in a car AND it is hot Addressing social concerns: The safety driver manager in a car AND car windows are closed. dialog and amount of “control” that he/she chooses to leave to the event managers. Complex situations could include. interface to select options and choices about the level of another car approaching. GPS. “How do I get to Spring Valley Rd?” A general approach to privacy. Driver preferences: typical routes. and into the service as input various media (audio. the car may be in New York State already (that is defined via GPS) and it may be known that the 7. through the use of IBM WebSphere Personalization complex or abstract. it may well learning module that detects and records customer deal with such complex situations as the following: feedback.

). One can also offer reduced insurance payments to Maison. For example. of the ICASSP 2001 system to address privacy concerns that drivers may [4] S. Liam Comerford. devices and drowsiness can be addressed by a special [6] P. some one who is more worried about young drivers may request that this information 10. Another way is that a driver 9. revenues for Telematics. recognition and automotive applications: using speech to perform in-vehicle tasks. [2] K.) without compromising a . chronic illnesses that can affect driving – like diabetes) downloading music titles etc. laughing. Gopalakrsihnan P. Providing new services in a car in a car. interrupt a dialog between the driver and a car system if. This interface requires interactions D. physical conditions (tired. "Performance of the IBM Large number. user interfaces in Telematics will have a significant Driver overall readiness for safe driving can be impact on driver safety and they will be the subject of evaluated by a safety manager in his/her car. N'adas and speech interface. Bain. drowsy. We showed that basic speech IEEE Transactions on Information Theory. interested in this approach since it allows reduction in [5] Gellatly. There is. We expect that novel ideas in age). sick. Conclusion Factors and Ergonomics Society 42nd Annual Meeting. an acknowledged need to reduction of conventional speech processes to low – provide a more dynamic arrangement to highlight a resources processing was done by reducing a signal variety of potentially dangerous situations to drivers of processing and decoding load in such a way that it did other cars and to ensure that drivers of other cars do not not significantly affect decoding accuracy and by the bear the added responsibility of discovering this development of quasi-NLU principles. upset) Safety Driver Manager would allow use of various . Successful implementation of . The her car. "An inequality for rational functions with with workload. interactions can be done in a low-resource embedded January 1991 processor and this allows a development of a useful 6 . This allows the Assistant”.announce this fact on the back of the window of his or local component of Safety Driver Manager. Faisman.R. Gopalakrishnan." other modules. Nahamoo. and Mahesh method allows individual car drivers to customize their Viswanatha for useful discussions.. have. R. We would like to thank manager in one car has access to information to a driver Sara Basson. calm.. pp 41-44. Basson. S. 37(1). as this presents its own that an important application like Artificial Passenger risks. attention (looking on a road or navigation map driver safety. Deligne.g.W. Human Factors and related to a driver safety. If a driver safety manager finds that there is Vocabulary Continuous Speech Recognition System on additional risk associated with driver behavior in a car the ARPA Wall Street Journal Task. Special Issue on Accessibility” driver safety manager in another car and this manager [3] Comerford L. We observed themselves through observation. In the paper we suggested that such important issues (Santa Monica CA. driver profile (number of traffic accidents. event. Olsen.A. References be provided to his/her driver safety manager rather than an overall estimation of risk expressed as a single [1] L.1.. and Dingus. Sally Lake. Eide. and Sedivy J. “The IBM personal Speech surrounding environment for this car. cars. D. this paper regarding the use of speech and distributive There can be several ways to assess this information . Acknowledgements. Proc. Kanevsky. Kanevsky It can also advise someone who is calling a driver that “LIBERATED LEARNING and IBM ViaScribe: that driver is busy and should not be disturbed at this Accessibility through speech recognition submitted to time.. Frank D. Information about driver conditions can be can be sufficiently entertaining for a driver with provided from sensors that are located in that car. factors. Employers of fleet tracks may be particularly Proceedings of EUROSPEECH 01. 1247-1251. The information can be sent anonymously to the IBM Systems Journal. then. listening to e-mail) comfortable and it can be a significant source of . would then adjust the risk factor in its estimation of the Gopinath R. D. and this evaluation can then be sent to the driver safety managers in nearby cars. Ponani Gopalakrishan. talking over environment is important to make the driver telephone. It can be intensive research and development in forthcoming ranked by some metrics (e. located nearby. T. such as controlling Telematics Ergonomics Society. E. P.. A.. profile and from sensors in other cars. talking to a baby in a back sit. A. has services in cars (like reading e-mail. H. D. dialog.S. on a scale from 1 to 5 ) years at IBM and other laboratories. Speech traffic accidents. B. 2001. Printz. navigation. The relatively little dialog complexity requirements – following are examples of the information about drivers playing simple voice games with a vocabulary that can affect driver conditions: containing a few words." ICASSP 1995. mood (angry. Bahl et al. privacy. advice and help in priorities and use personal estimators for driving risks preparation of the paper. Sedivy: “Low-resource a driver if s/he agrees to disclose information to other speech recognition of 500-word vocabularies”. Proceedings of the Human 8. situation and applications to some statistical estimation problems. it may prevent a telephone ringing or vol. Paul Moskowitz. This second Steve Ross. Kanevsky. 1998).. A. Gopinath. J.

and Zadrozny. Some elements of this approach (e. E. U.S. 6. and Yashchin..902. “Language model adaptation via network of similar users”.visualexpert.. 2002. published in International Journal of Speech Technology. Insurance) [14] Marvin C. Campbell. [12] Kanevsky. DOT-HS-034-3-535-77 (TAC). W. R.505. Patent No. J. Patent No. Report No. 2003. N. Montreal. 2003. Shinar. Ronald Perez. S.com/Resources/roadaccidents.... N 1. (1977) Tri-level study of the causes of traffic accidents. Privacy Manager.. Zlatsin.136 . 2004. Sedivy. C. Patents: No. [11] Kanevsky. J. “Speaker model adaptation via network of similar users”. Richman and James L.529. N. Tumbas.S. “Sleep prevention dialog based car system”. D. 6. and Zadrozny. Wolf.[7] Marc Green & John Senders. [16] Treat. Maximum Mutual Information Estimation and the Speech Recognition Problem". L. Patent No.208. V. v.7. W. 6. D. Hume. D. "Hidden Markov Models. R. and Castellan. A. 6.442.S.519. Jung-Mu Tang. U..968 . 2004.. [10] Kanevsky. 7 . McCallum. John L. D. 2002. [9] Kanevsky. [13] Paul Moskowtiz.html [8] Kanevsky.). “Educational monitoring method and system for improving interactive skills based on participants on the network”. U. McGill University. D. [15] Y.. 1991. "Speech Recognition and In-Vehicle Telematics Devices: Potential Reductions in Driver Distraction". U. www. R.S. McDonald. W. C. S. E. Brown.236.S.g.484.. T. D. Normandin.. “Method and system for off-line detection of textual topical changes and topic identification via likelihood based methods for improved language modeling”. and Zadrozny.. Libal. J. “Mobile Networks and Applications”s Volume 9 issue 5. and Pickover. R.. Patents: No 6. Joel B.. PhD thesis. Stanisfer. U.. 2001. Mayer. Moninder Singh.. D.

8 .