cobotDS User Manual* Version 1.0b
Contents
1. 0 cobot Hi who is this 1a PhoneUser Phone user 1b cobot Secret password 2a PhoneUser Five one two seven 2b cobot Hi PhoneUser What would you like to do Here the utterances of the human PhoneUser are of course spoken to cobot over the phone while the utterances of cobot are spoken using text to speech synthesized speech At this point cobot has accepted the password of PhoneUser and is ready to serve Meanwhile back in the Living Room of LambdaMOO users will have seen an announcement from cobot that he has received a phone call as in the following sample transcript again we introduce numerical labels on the acts which are not intended to be syn chronized with the phone conversation labels We are still tinkering with the exact wording of the phone related prompts to both the phone user and LambdaMOO users so don t be surprised if the actual system differs slightly 1 Buster to HFh We should definitely add the script of the Coen brothers Unabomber biopic 2 HFh to Buster Cool dat Also the Blair Witch JonBenet thing w 4 Buster nods to HFh At this point users in the LambdaMOO Living Room have been made aware that PhoneUser is on the phone with cobot line 3 and may expect cobot to soon be conveying messages from PhoneUser into the Living Room We return to PhoneUser who now engages cobot in a series of requests and queries 3a PhoneUser Wave 3b cobot You wave 4a Pho2. command fails is the Cliche grammar 6 4 Personal Grammars The last grammars that we are providing in our initial system are those provided by the phone users themselves Despite our best attempts to pro vide some basic coverage of common useful utterances through the Cliche and Small Talk grammars we will inevitably miss many particular phrases that individual users may want to be able to pass using the say command We thus invite each phone user to provide us with an evolving list of such personalized phrases When a phone user successfully invokes the com mand grammar personal cobot will set the active grammar to be the personal grammar provided by the particular user as identified during lo gin to cobot We will solicit source material for each personal grammar from phone users This will typically be provided as a simple text file of English phrases words and sentences all lower case and one to a line We will take this source material and compile it as a grammar after which users should be able to pass utterances from the source using the say command with active grammar set to personal The mechanism by which users can update and re place their personal grammars are still to be determined as is the frequency with which we will perform such updates updating grammars requires both a compilation process and for cobot to be taken down We will keep our phone users informed of the necessary details but it is safe to expect that
3. us know The contacts for cobotDS are Charles Isbell Lamb daMOO user HFh RL email isbell research att com and Michael Kearns LambdaMOO user Buster RL email mkearns research att com 13
4. user needs time to contemplate their next action or is awaiting a response to an earlier verb pushed to LambdaMOO During such periods it is inconvenient to have the main command loop continually prompting the phone user for the next action The sleep command is designed to force cobot to idle at the phone user s command After uttering the sleep command cobot will announce that he is now sleeping and a silent period will ensue During this period cobot will not prompt the phone user for additional commands but will continue to announce the arrival of verbs pushed from LambdaMOO At any point during the silent period the phone user may wake cobot by simply uttering their next command The help command will eventually provide on line audio help for phone users but is not implemented yet The phone user should utter the quit command to terminate a call to cobot Please make your best effort to end calls explicitly with quit rather than simply hanging up in the middle of a dialogue 10 Appendix 10 1 Listing of Phone Commands The following verbs from the LambdaMOO social core are currently avail able as cobot phone commands blush bow cackle chuckle comfort cringe cry french giggle grin hug kiss laugh nod poke say shrug sigh smile smirk wave whisper wink yawn All accept an optional user name argument As already discussed the say command will cause cobot to prompt the user for an utterance In addition to
5. HFh have invoked new functionality on cobot lines 10 11 and 14 in order to return PhoneUser s emotes and also to pass an utterance to PhoneUser Thus PhoneUser would now hear something like Sa 8b 9a 9b 10a 10b 11a 11b PhoneUser Cackle GreenBean cobot You cackle to Greenbean You got a hug from Buster PhoneUser Grin Natto cobot You grin to Natto You got a smile from HFh Buster says How did your interview go to you PhoneUser Say Buster cobot What message would you like to pass PhoneUser No point crying over spilt milk cobot You say No point crying over spilt milk to Buster Shortly afterwards cobot will announce PhoneUser s cackle grin and utterance to Buster in the Living Room This interleaved example conveys the spirit and basic functionality of the system we are building In the remainder of this document we provide a detailed guide for the use of cobot for the phone user We emphasize that we have implemented and describe here only the most basic features on this service and will be rolling out many others over the months to come Again we invite suggestions 3 Talking to the Phone User from LambdaMOO Before we describe how to use cobot as a phone user we briefly explain how any LambdaMOO user can pass a verb or message to the phone user through cobot There are two separate cases both invoked by directed speech at cobot via cobot 7 The syntax of the first instance i
6. and adding common phrases as we think of them and solicit suggestions for additions The Small Talk grammar contains staples such as How are you Hello I am fine Yes I am at home and many others A complete listing of the elements of the Small Talk grammar along with a simple Perl script for searching the grammar by some indexed categories such as greeting phrases basic responses phrases about domestic or work life etc will be provided to phone users in a separate document Again any time the Small Talk grammar is active all say utterances will be recognized using the Small Talk grammar only 6 3 The grammar Command We now have two different grammars Cliche and Small Talk that can be applied to the utterance of a say command How can the phone user choose which grammar should be activated on any given say The answer is the grammar command which can be invoked at the main command loop by simply saying grammar cobot will then prompt the phone user for a gram mar name which currently must be one of the utterances cliche small talk or personal The successful invocation of the grammar command will result in the active grammar being set to the designated argument This grammar will be applied to all subsequent say commands until the grammar is changed again or the call terminates Again the default grammar active at the start of every call or made active any time a grammar
7. cobotDS User Manual Version 1 0b September 22 2000 1 Introduction Many thanks for agreeing to be one of the first users of cobot s new dialogue system cobotDS We appreciate your participation and will take your com ments and suggestions seriously The design of any spoken dialogue system is a complex and difficult task that usually takes many iterations and lots of trial and error Because cobotDS will tackle a number of issues that have not been faced before in dialogue systems research your help could prove especially valuable and influential in the direction we take and the solutions we adopt We also hope you will find using cobotDS interesting and fun What exactly is cobotDS In brief cobotDS is a computerized telephone system that uses speech recognition text to speech and artificial intelligence technology It allows users to call cobot from any domestic US telephone on a toll free number interact with cobot over the phone using spoken English and have cobot relay messages and information back and forth between the phone user and LambdaMOO Using cobotDS you will be able to interact with the LambdaMOO envi ronment and users over the phone in many of the same ways you already do when in the MOO itself You will be able to wave grin poke hug etc from the phone to a room in general or to a specific user and LambdaMOO users will be able to pass such emotes to the phone user as well The phone user will also be able to pas
8. ge and whisper may also be invoked with the analogous syntax The use of such silent direct speech verbs will obviously prevent the passed verb from being publicly announced in LambdaMOO 3Note that such acknowledgements did not appear in the LambdaMOO transcripts pro vided in the example of the Introduction since these transcripts were from the viewpoint of an unseen user on Using Speech Recognition Most importantly when giving your pass word please speak each digit individually Thus the password 5127 must be spoken five one two seven and not variants such as fifty one twenty seven or five thousand one hundred and twenty seven The digit 0 must be pronounced as zero and not oh Following a successful login which is attempted a few times before cobot gives up and ends the call cobot will prompt you with a greeting confirming your user name and then ask you what you want to do At this point you are in the main command loop of cobot 5 The Main Phone Command Loop Under normal use the core of a successful call to cobot will consist of a series of dialogue turns alternating between utterances by the phone user and cobot In a typical turn the phone user makes a request of some sort to cobot and in response cobot confirms the command and may also provide additional information such as recent emotes or messages for the phone user from LambdaMOO The commands available to the phone user a
9. mand as it is used in LambdaMOO is meant to be a powerful general purpose verb for social communication and we would like to recapture at least some of this power for the cobot phone user Our approach is to make several different grammars available for use with the say command and to allow the phone user to choose which of these is active during any given invocation of say We have initially implemented two grammars and will allow phone users to provide us with a third personal grammar that can be modified by them periodically We now describe each of these grammars in turn 4 6 1 The Cliche Grammar The Cliche grammar is the default for the say command and will always be active during a say unless another grammar is explicitly set by the phone user via the grammar command described shortly The Cliche grammar consists of approximately 3000 common English cliches sayings and phrases A full listing of the items in the Cliche grammar will be provided to phone users in a separate document When the Cliche grammar is active the phone user invoking the say command will be able to utter an item in the grammar whereupon it will hopefully be recognized successfully and passed on to LambdaMOO by cobot Although a phone user s say utterances using exclusively the Cliche grammar would be highly stylized the grammar is rich enough to contain ap propriate remarks and responses for many conversational situations Becker reference Howeve
10. neUser Where 4b cobot I am in the Living Room 5a PhoneUser Who 5b cobot I am with Buster HFh Betwixt Natto and GreenBean 6a PhoneUser Hug Buster 6b cobot You hug to Buster 7a PhoneUser Poke HFh 7b cobot You poke to HFh In 3a PhoneUser waves to all those present in cobot s current location the Living Room In turns 4a b and 5a b PhoneUser asks cobot where he is and what other users are there and receives the answers from cobot These two turns being exchanges of information entirely between PhoneUser and cobot have no visible consequences for the Living Room occupants However the acts by PhoneUser in 3a 6a and 7a might appear to the Living Room as follows 5 cobot holds up a sign that reads PhoneUser passes on a wave from the phone 6 Betwixt to HFh The scariest part of Blair Witch JonBenet was the mother cobot exclaims PhoneUser has just called me on the phone 00 10 11 12 13 14 cobot to Buster This is from PhoneUser cobot hugs Buster warmly HFh to Betwixt Really It was the tiara collection scene I couldn t watch Buster to cobot phone hug Buster to cobot phone say How did your interview go cobot pokes HFh cobot to HFh That was from PhoneUser HFh to cobot phone smile Here we see that the emotes dictated by PhoneUser to cobot have now been announced by cobot in the Living Room lines 5 7 8 12 and 13 Further Buster and
11. nounce ment of the resulting act is quite analogous as well However note that the phone user s verbs will always be conveyed to LambdaMOO via an an nouncement by cobot and thus will not appear as if they were processed by the LambdaMOO server Simply put the metaphor is that the phone user is actually chatting with cobot on the phone with cobot keeping LambdaMOO informed about the activity in the call where appropriate 6 The say Command Speech Recognition and Grammars An important exception to the basic command invocation on cobot described above is the say command To begin with when the phone user invokes the say command by speaking for example either say or say Buster or some other user name cobot will then prompt the user for the message they would like to pass to LambdaMOO Following this prompt the user should simply speak the English sentence they would like conveyed subject to some important and nontrivial constraints Here is where some background on speech recognition is required The goal of speech recognition is to automatically compute an accurate text transcription of a spoken utterance Highly accurate speaker independent speech recognition of arbitrary natural language utterances is beyond cur rent technology Instead all speech recognizers make use of some underlying grammar describing which utterances are allowed by the recognizer For our current purposes we may think of this grammar as simply an e
12. phone users will be able to modify their grammars roughly weekly but not daily The use of personal grammars is one of the more interesting research aspects of cobot and we invite creativity by phone users in choosing their source material For example some users may want to hand craft a lengthy slowly changing list of favorite utterances and sayings not covered by the other grammars or to create specialized functional phrases appropriate to the details of their life I am at Wanda s house I am playing Ultimate 10 today Others may want to manage their grammar more dynamically anticipating and swapping in phrases that will be useful only for a short time following the grammar update I am calling from my vacation in Wyoming Still others may want to experiment with adding text from favorite books movies poems and so on 7 Hints on Using Speech Recognition As you might have guessed from the last section speech recognition is an imperfect technology Performance can vary widely depending on many fac tors the particular utterance the particular speaker the particular phone called from Here are some hints for trying to get the best performance e Try to minimize the amount of acoustic noise in your calls Calls made from a quiet room or phone booth on a land line phone are likely to do better than calls made from a cell phone at a busy intersection Nevertheless feel free to try it anywhere we are interes
13. r there are several difficulties in relying solely on the Cliche grammar to provide the phone user with a rich set of utterances 4We emphasize that these grammars are simply our starting point and that we are open to suggestions for additional grammar sources and types e Searching for the right phrase in such a large grammar in real time during a live conversation while on the phone is difficult e Being a list of cliches and common sayings the grammar is missing some obvious conversational necessities such as basic greetings ques tions and replies How are you I am at work yes no etc e The grammar is also missing particular utterances that a particular phone user might find useful either repeatedly or even for short periods of time e Even though the Cliche grammar is a highly restricted fragment of the English language phone users will still find the speech recognition sometimes inaccurate For these reasons we view the Cliche grammar as serving a mixture of practical and entertainment purposes not a general purpose conversational mechanism We attempt to partially address the second and third items in the list above by the introduction of the Small Talk and Personal grammars 6 2 The Small Talk Grammar The Small Talk grammar is meant to provide a more basic and coherent set of useful utterances for common conversation than is provided by the highly stylized Cliche grammar We are building the Small Talk grammar by h
14. re of the form lt command gt lt user name gt where the command is a required utterance and the user name argument is optional or absent With a few exceptions discussed later all of the choices for the command are standard LambdaMOO social core verbs such as wave smile cackle and so on A list of the verbs currently supported is provided in the Appendix Until we have the chance to develop the speech recognition tools to handle a large set of LambdaMOO user names the allowed user names are restricted to a small group that is listed in the Appendix However the phone user can still emote and speak to the LambdaMOO room in general and the context of conversation so far may be sufficient for LambdaMOO users to recognize phone user responses to their acts As shown in the example in the Introduction the phone user might typically invoke a sequence of commands on cobot such as wave buster smile and cackle In each case the command will cause cobot to announce the phone user s actions publicly in whatever room of LambdaMOO cobot currently happens to occupy When the requested command has an allowed user name as an argument cobot will announce that the verb is invoked on that user while if a user name is absent cobot will announce that the phone user has invoked the verb to everyone in the room Thus spoken verb invocation on cobot has essentially the same syntax as it has in the textual environment of LambdaMOO and the an
15. s cobot phone lt verb gt Here phone is a required field telling cobot that the following text is intended for the current phone user and the verb is any verb such as hug wave etc cobot will privately acknowledge the LambdaMOO user s command and momentarily the phone user will hear the announcement of the verb from cobot along with which LambdaMOO user passed the verb In the second instance the syntax is cobot phone say lt utterance gt In this case the utterance is an arbitrary English sentence word or phrase The utterance will be read to the phone user by cobot along with the identity of the LambdaMOO user sending it LambdaMOO users should remember that the phone user must process the message over a telephone in real time and should be careful not to overload the phone user with verbs and messages 4 Calling cobot on the Phone Your phone user name for cobot will be the same as your LambdaMOO user name In addition we will provide you with a four digit password code for phone access to cobot please keep this code private We will also provide you with cobot s telephone number This number is a toll free call from within the United States Upon calling cobot you will be prompted for your user name and pass word as in the example given in the Introduction Please speak your re sponses to these in a clear and natural manner See also Section 7 Hints Other directed speech verbs such as pa
16. s certain restricted but hopefully sufficient natural language utterances to LambdaMOO and LambdaMOO users will be able to pass arbitrary utterances to the phone user Additionally cobotDS will cobotDS is a system under active development and will be undergoing rapid growth and change We will periodically revise this document accordingly provide a number of informational commands for the phone user that al low them to find out what is going on in LambdaMOO and will eventually replicate at least some of the existing functionality of cobot social statistics such as verb o meter commands general chat functionality etc Thus cobotDS will provide rich multiparty interaction between a phone user cobot and the denizens of LambdaMOO We anticipate that in the early stages of development the primary use of cobotDS will be to commu nicate via phone with other human LambdaMOO users but eventually some phone users may call to interact primarily or partially with cobot himself 2 A Sample Session Let s make this a little more concrete by examining both sides phone user and LambdaMOO of a typical interaction with cobot Imagine that cobot is going about his business as usual in the Living Room and a person such as yourself with LambdaMOO user name PhoneUser calls cobot using the toll free number The following initial interaction would be typical where we have introduced numerical labels on the turns for ease of exposition
17. ted in the system being used over a range of conditions and environments just try not to get frustrated e Try to speak clearly and articulate your words carefully without being unnatural In particular speaking deliberately slowly and drawing out the syllables may result in worse performance rather than better Nevertheless do experiment around to see what seems to give the best performance for you 8 Informational Commands where and who The commands where and who which take no arguments will cause cobot to report to the phone user which room of LambdaMOO cobot currently occupies and which other LambdaMOO users are also in that room re spectively 9 Other Commands repeat sleep help and quit Other commands available at the main command loop of cobot include re peat sleep help and quit none of which takes any arguments The repeat command will simply cause cobot to repeat the last prompt to the phone user including any emotes or messages passed to the phone 11 user from LambdaMOO Since the synthesized speech of arbitrary messages passed from LambdaMOO can sometimes be difficult to understand on first hearing the repeat command allows the phone user to replay such messages before moving on to their next command Note that past prompts and messages are lost once a non repeat command is given Later versions of cobot may provide more elaborate audio scrollback functionality There may be times when the phone
18. the social core verbs cobot also recognizes the no argument commands grammar help repeat sleep where who quit The grammar command will cause cobot to prompt for a grammar name 12 10 2 Current LambdaMOO Phone User List The current list of LambdaMOO users with a cobot phone user account are AcidHorse Betwixt Boiledhead Buster CrashLander Cyberat Gabaldon Goat GreenBean HFh Huey LuciaLibra Milly Natto and Oedipus LuciaLibra should be refered to as Lucia due to trouble we had with the speech recognizer on this name These are also the only LambdaMOO user names that cobot will recognize as arguments to the social core verbs 10 3 Privacy Policy The cobotDS system will of course keep system logs that include your di alogues with the system and the phone number you call from All of this information will be kept absolutely confidential and will not be accessible to anyone outside of our small research group of approximately five people all of whom are directly working on the project In general we will not use this information for anything other than debugging and improving the system 10 4 Contacting Us We solicit your feedback and suggestions on the use functionality and design of cobotDS and invite you to keep in touch with us both in LambdaMOO and via email We have monitoring tools that will automatically inform us when cobotDS is not working but if you notice anything especially anoma lous please let
19. xhaus tive listing of all those words phrases and sentences that the recognizer will attempt to recognize Any utterance not contained in this list should not be expected to be recognized A typical dialogue system cobot s included ac tually uses several different grammars each appropriate in different dialogue contexts Why not just set this grammar to be the entire English dictionary or a phenomenally large corpus of English sentences The main reason is that there is a trade off between the size or complexity of the grammar and the accuracy of recognition Simply put a very small grammar will result in highly accurate recognition but only of those items appearing in this lim ited grammar while a very large grammar may in principle permit more expressive utterances but with frequently poor accuracy words being mis recognized dropped or added For this reason most applications of speech recognition attempt to balance the burden placed on users by forcing them to use restrictive language with the burden placed on the speech recognizer by having to cope with many possible utterances In many cases this trade off has a nice intermediate solution For example the grammar we have written for the main command loop of cobot is mercifully small and simple since it exploits the fact that there are rather few verbs in the social core and we have artificially limited the number of possible user names for now However the say com
Download Pdf Manuals
Related Search