This document introduces the W3C Multimodal Interaction Framework, and identifies the major components for multimodal systems. Each component represents a set of related functions. The framework identifies the markup languages used to describe information required by components and for data flowing among components. The W3C Multimodal Interaction Framework describes input and output modes widely used today and can be extended to include additional modes of user input and output as they become available.
This section describes the status of this document at the time of its publication. Other documents may supersede this document. The latest status of this document series is maintained at the W3C.
W3C's Multimodal Interaction Activity is developing specifications for extending the Web to support multiple modes of interaction. This document introduces a functional framework for multimodal interaction and is intended to provide a context for the specifications that comprise the W3C Multimodal Interaction Framework.
This document has been produced as part of the W3C Multimodal Interaction Activity, following the procedures set out for the W3C Process . The authors of this document are members of the Multimodal Interaction Working Group (W3C Members only). This is a Royalty Free Working Group, as described in W3C's Current Patent Practice NOTE. Working Group participants are required to provide patent disclosures.
Please send comments about this document to the public mailing list: firstname.lastname@example.org (public archives). To subscribe, send an email to <www-multimodal-request@w3. org> with the word subscribe in the subject line (include the word unsubscribe if you want to unsubscribe).
A list of current W3C Recommendations and other technical documents including Working Drafts and Notes can be found at http://www.w3.org/TR/.
The purpose of the W3C multimodal interaction framework is to identify and relate markup languages for multimodal interaction systems. The framework identifies the major components for every multimodal system. Each component represents a set of related functions. The framework identifies the markup languages used to describe information required by components and for data flowing among components.
The W3C Multimodal Interaction Framework describes input and output modes widely used today and can be extended to include additional modes of user input and output as they become available.
The multimodal interaction framework is not an architecture. The multimodal interaction framework is a level of abstraction above an architecture. An architecture indicates how components are allocated to hardware devices and the communication system enabling the hardware devices to communicate with each other. The W3C Multimodal Interaction Framework does not describe either how components are allocated to hardware devices or how the communication system enables the hardware devices to communicate. See Section 6 for descriptions of several example architectures consistent with the W3C multimodal interaction framework.
This document does not discuss two important multimodal system architecture variations: distributed multimodal systems and multi-user multimodal systems. These topics are under discussion by the working group,
Figure 1 illustrates the basic components of the W3C multimodal interaction framework.
Human user--A user who enters input into the system and observes and hears information presented by the system. In this document, we will use the term "user" to refer to a human user. However, an automated user may replace the human user for testing purposes. For example, an automated "testing harness" may replace human users for regression testing to verify that changes to one component do not affect the user interface negatively.
Input — An interactive multimodal implementation will use multiple input modes such as audio, speech, handwriting, and keyboarding, and other input modes. The various modes of input will be described in Section 3.
Output — An interactive multimodal implementation will use one or more modes of output, such as speech, text, graphics, audio files, and animation. The various modes of output will be described in Section 4.
Interaction manager — The interaction manager controls the sequence of exchange for information between the human user and the application functions. It may support a variety of interaction styles including:
The interaction manager may use (1) inputs form the user, (2) the session context, (3) external knowledge sources, and (4) disambiguation, correction, and configuration subdialogs to determine the user's focus and intent. Based on the user's focus and intent, the interaction manager also (1) maintains the context and state of the application, (2) manages the composition of inputs and synchronization across modes, (3) interfaces with business logic, and (4) produces output for presentation to the user. One way to implement a strategy for determining the user's focus and intent is a interaction script. A interaction script may be expressed in languages such as Tsuneo Nitta's XISL , or languages used for the control of user interfaces such as XHTML+Voice Profile or SALT . The interaction manager may make use of context and history to interpret the inputs. In some architectures, the interaction manager may be distributed across multiple components, utilizing an event based mechanism for coordination.
Application functions — Several application-specific functions provide database access, transaction processing, and application-specific calculations. The exact format of the interface to application functions is dependent upon the application and is outside the scope of this document.
Figure 2 illustrates the various types of components within the input component.
Recognition component — Captures natural input from the user and translates the input into a form useful for later processing. The recognition component may use a grammar described by a grammar markup language. Example recognition components include:
Other input recognition components may include vision, sign language, DTMF, biometrics, tactile input, speaker verification, handwritten identification, and other input modes yet to be invented.
Interpretation component — May further process the results of recognition components. Each interpretation component identifies the "meaning" or "semantics" intended by the user. For example, many words that users utter such as "yes," "affirmative," "sure," and "I agree," could be represented as "yes."
Integration component — Combines the output from several interpretation components
Some or all of the functionality of this component could be implemented as part of the recognition, interpretation, or interaction components. For example, audio-visual speech recognition may integrate lip movement recognition and speech recognition as part of a lip reading component, as part of the speech recognition component, or integrated within a separate integration component. As another example, the two input modes of speaking and pointing are used in
"put that," (point to an object), "there," (point to a location)
and may be integrated within a separate integration component or may be integrated within the interaction manager component.
Either the user or the system may create information that may be routed directly to the interaction manager without being encoded in EMMA. For example, audio is recorded for later replay or a sequence of keystrokes is captured during the creation of a macro.
Information generated by other system components may be integrated with user input by the integration component. For example, a GPS system generates the current location of the user, or a banking application generates an overdraft to prohibit the user from making additional purchases.
The output for each interpretation component may be expressed using EMMA, a language for representing the semantics or meaning of data.
Figure 3 illustrates the components within the output component.
Generation component — The generation component determines which output mode or modes will be used for presenting information from the interaction manager to the user. The generation component may select a single output mode or it may select complementary or supplementary modes. The "internal representation" language used to describe the output from the generation component is under discussion by the working group.
Information from the interaction manager may be routed directly to the appropriate rendering device without being encoded in an internal representation. For example, recorded audio is send directly to the audio system.
Styling component — This component adds information about how the information is "layed out." For example, the styling component for a display specifies how graphical objects are positioned on a canvas, while the styling component for audio may insert pauses and voice inflections into text which will be rendered by a speech synthesizer. Cascading Style Sheets (CSS) could be used to modify voice output.
Rendering component — The rendering component converts the information from the styling component into a format that is easily understood by the user. For example, a graphics rendering component rectangle displays a vector of points as a curved line, and a speech synthesis system converts text into synthesized voice.
Each of the output modes has both a styling and rendering component.
The voice styling component constructs text strings containing Speech Synthesis Markup Language tags describing how the words should be pronounced. This is converted to voice by the voice rendering component. The voice styling component may also select prerecorded audio files for replay by the voice rendering component.
The graphics styling component creates XHTML, XHTML Basic, or SVG markup tags describing how the graphics should be rendered. The graphics rendering component converts the output from the graphics styling component into graphics displayed to the user.
Other pairs of styling and rendering components are possible for other output modes. SMIL may be used for coordinated multimedia output.
To illustrate the component markup languages of the W3C Multimodal Interaction Framework, consider this simple use case. The human user points to a position on a displayed map and speaks: "What is the name of this place?" The multimodal interaction system responds by speaking "Lake Wobegon, Minnesota" and displays the text "Lake Wobegon, Minnesota" on the map. The following summarizes the actions of the relevant components of the W3C Multimodal Interaction Framework:
Human user — Points to a position on a map and says, "What is the name of this place?"
Speech recognition component — Recognizes the words "What is the name of this place?"
Pointer recognition component — Recognizes the x-y coordinates of the position to which the user pointed on a map.
Speech understanding component — Converts the words "What is the name of this place?" into an internal notation.
Pointing understanding component — Converts the x-y coordinates of the position to which the user pointed into an internal notation.
Integration component — Integrates the internal notation for the words "What is the name of this place?" with the internal notation for the x-y coordinates.
Interaction manager component — Converts the request to a database request, submits the request to a database management system which returns the value of "Lake Wobegon, Minnesota". The interaction manager converts the response into an internal notation and sends the response to the generation component.
Generation component — Decides to present the result as two complementary modes, voice and graphics. The generation component sends internal notation representing "Lake Wobegon, Minnesota" to the voice styling component, and sends internal notation representing the location of Lake Wobegon, Minnesota on a map to the graphics styling component.
Voice styling component — Converts the internal notation representing "Lake Wobegon, Minnesota" into SSML.
Graphics styling component — Converts the internal notation representing the "Lake Wobegon, Minnesota" location on a map into HTML notation.
Voice rendering component: Converts the SSML notation into acoustic voice for the user to hear.
Graphics styling component: Converts the HTML notation into visual graphics for the user to see.
There are many possible multimodal architectures that are consistent with the W3C multimodal interaction framework. These multimodal architectures have the following properties:
Property 1. THE MULTIMODAL ARCHITECTURE CONTAINS A SUBSET OF THE COMPONENTS OF THE W3C MULTIMODAL INTERACTION FRAMEWORK. A multimedia architecture contains two or more output modes. A multimodal architecture contains two or more input modes.
Property 2. COMPONENTS MAY BE PARTITIONED AND COMBINED. The functions within a component may be partitioned into several modules within the architecture, and the functions within two or more components may be combined into a single module within the architecture.Property 3. THE COMPONENTS ARE ALLOCATED TO HARDWARE DEVICES. If all components are allocated to the same hardware device, the architecture is said to be centralized architecture . For example, a PC containing all of the selected components has a centralized architecture. A client-server architecture consists of two types of devices, several client devices containing many of the input and output components, and the server which contains the remaining components. A distributed architecture consists of multiple types of devices connected by a communication system.
Property 4. THE COMMUNICATION SYSTEMS ARE SPECIFIED. Designers specify the protocols for exchanging messages among hardware devices.
Property 5. THE DIALOG MODEL IS SPECIFIED. Designers specify how modules are invoked and terminated, and how they interpret input to produce output.
The following examples illustrate architectures that conform to the W3C multimodal interaction framework.
Example 1: Driving Example (Figure 4)
In this example, the user wants to go to a specific address from his current location and while driving wants to take a detour to a local restaurant (The user does not know the restaurant address nor the name). The user initiates service via a button on his steering wheel and interacts with the system via the touch screen and speech.
Property 1. The driving architecture contains the components illustrated in Figure 4: a graphical display, map database, voice and touch input, speech output, local ASR, TTS Processing and GPS.
Property 2. No components are partitioned or combined with the possible exception of the integration and interaction manager components, and the generation and interaction components. There are two possible configurations, depending upon whether the integration component is stand alone or combined with the interaction manager component:
Information entered by the user may be encoded into EMMA (Extensible MultiModal Annotation Markup Language, formerly known as the Natural Language Semantic Markup Language) and combined by an integration component (shown within the dotted rectangle in Figure 4) which is separate from the interaction manager.
Information entered by the user may be recognized and interpreted and then routed directly to the interaction manager, which performs its own integration of user information
There are two possible configurations, depending upon whether the generation component is stand alone or combined with the interaction manager component:
Information from the interaction manager may be routed to the generation component, where multiple modes of output are generated and the appropriate synchronization control created.
Information may be be routed directly to the styling components and then on to the rendering components. In this case, the interaction manager does its own generation and synchronization.
Property 3. All components are allocated to a single client side
hardware device onboard the car. In Figure 4, the client is
illustrated by a pink box containing all of the components.
Property 4. No communication system is required in this centralized architecture.
Property 5. Dialog Model: The user wants to go to a specific
address from his current location and while driving wants to take a
detour to a local restaurant . (The user does not know the
restaurant name or address.) The user initiates service via a
button on his steering wheel and interacts with the system via the
touch screen and speech.
Example 2: Name dialing (Figure 5)
The Name dialing example enables a user to initiate a call by saying the name of the person to be contacted. Visual and spoken dialogs are used to narrow the selection, and to allow an exchange of multimedia messages if the called person is unavailable. Call handling is determined by a script provided by the called person. The example supports the use of a combination of local and remote speech recognition.
Property 1: The architecture contains a subset of the components of the W3C Multimodal Interface Framework.
Property 2: No components have been partitioned or combined with the possible exception of the integration component and interaction component, and the generation component and the interaction component (as discussed in example 2).
Property 3. The components in pink are allocated to the client and the components in green are allocated to the server. Note that the speech recognition and interpretation components are on both client and server. The local ASR recognizes basic control commands based upon the ETSI DES/HF-00021 standardized command and control vocabulary, and the remote ASR recognizes names of individuals the user wishes to dial. (The vocabulary of names is too large to maintain on the client, so it is maintained on the server.)
Property 4. Communications system is SIP. SIP is a session initiation protocol and is a means for initiating communication sessions involving multiple devices, and for control signaling during such sessions.
Property 5. Navigational and control commands are recognized by the ASR on the client. When the user says "call John Smith," the ASR on the client recognizes the command "call" and transfers the following information ("John Smith") to the server for recognition. The application on the server then connects the user with John Smith's telephone.
Example 3: Form fill-in (Figure 6)
In the Form fill-in example, the user wants to make a flight
reservation with his mobile device while he is on the way to work.
The user initiates the service by means of making a phone call to
a multimodal service (telephone) or by selecting an application
(portal environment metaphor). The dialogue between the user
and the application is driven by a form-filling paradigm where the
user provides input to fields such as "Travel Origin:", "Travel
Destination:", "Leaving on date", "Returning on date". As the user
selects each field in the application to enter information, the
corresponding input constraints are activated to drive the
recognition and interpretation of the user input.
Property 1: The architecture contains a subset of the components of the W3C Multimodal Interface Framework, including GPS and Ink.
Property 2: The speech recognition component has been partitioned into two components, one which will be placed on the client and the other on the server. The integration component and interaction component, and the generation component and the interaction component may be combined or left separate (as discussed in example 2).
Property 3. The components in pink are allocated to the client and the components in green are allocated to the server. Speech recognition is distributed between the client and the server, with the feature extraction on the client and the remaining speech recognition functions performed on the server.
Property 4. Communications system is SIP. SIP is a session initiation protocol and is a means for initiating communication sessions involving multiple devices, and for control signaling during such sessions.
Property 5. Dialog Model: The user wants to make a flight reservation with his mobile device while he is on the way to work. The user initiates the service via means of making a phone call to a multimodal service (telephone metaphor) or by selecting an application (portal environment metaphor). The dialogue between the user and the application is driven by a form-filling paradigm where the user provides input to fields such as "Travel Origin:", "Travel Destination:", "Leaving on date", "Returning on date". As the user selects each field in the application to enter information, the corresponding input constraints are activated to drive the recognition and interpretation of the user input. The capability of providing composite multimodal input is also examined, where input from multiple modalities is combined for the interpretation of the user's intent.