Using an Audio Companion Device

Noi Sukaviriya

Mobile Solutions/Personal Systems
IBM Thomas J Watson Research Center
noi@watson.ibm.com

CHI '97 Workshop: Speech User Interface Design Challenges

Short Biography

I have some experience building systems which coordinated (automatically generated) spoken text and animation in context-sensitive animated help. I played with a small command-based spoken interface, mostly to replace existing single commands using IN3. So far my experience has been mostly with visual interfaces. I have joined IBM and the mobile solution group for a year and a half. This group mainly looks at small devices hence the use of speech as an alternative interaction technique is strong for lack of screen space. I just started looking into speech interface design issues.

Design Context

The problem we are currently looking into is how an audio companion device could be used in our daily lives. This device is a small handheld device which users can carry with them all the time. The device has audio input and output capability, some limited amount of memory, and perhaps with or without a limited amount of a display surface. The goal is to give users connectivity to their back office systems or their home systems. The device could be connected wirelessly, but should always be operational in a disconnected mode.

The speech group at IBM has been working on dictation systems and speech input for command and control for quite some time now. The quality of the recognition has been acceptable and a commercial product has been released (VoiceType). Our group is looking to exploit the IBM speech recognition system for mobile use. One can think of many useful functions to be designed and incorporated into this device. Some of these functions will lead to using speech as a means to enter long input entries. For this workshop, I'll focus on the problem related to navigating and correcting long content entries using speech on this device.

Design Problem

Consider telling the device to send e-mail to Judy Smith and the content is "Judy, could you please pick up my children for me after work? I need to take my mom to the hospital this afternoon. I don't know how long it will take. Will call if the plan changes. Thank you." Imagine a user wanting to correct audio reports which he collected while sampling dishes at an International Food Contest. The report will be used later for his newspaper article about the contest. Consider a case where a user asks the system to read out a report halfly written while she was in the office and now she is on a long train ride to another city. She does not want to carry a heavy notebook system and wishes to utilize her time finishing up this report while sitting on the train.

Using dictation has the challenge on small devices since the dictation engine is currently too large to run locally. The device has to send the audio signal to a recognition server and receive the recognized text back for further access. The recognition may or may not be done while the user is speaking into the device depending on the status of the connection. If the user asks the system to play back what it has, first of all the feedback must be clear whether the input has been processed or not. When the device plays back the recorded message, there may be errors which the user wants to correct. Even if the recognition system is perfect, there will always be times when the user may want to correct input because wrong words were said or the user may wish to add or subtract to the text. The correction should be allowed whether the audio input has been recognized or not.

Regardless of the reasons for correction, the design problem has to do with positioning into the content of a single long input (an e-mail message, the content of a report, annotating an audio textbook, preparing a lecture, etc.), selecting the range of the input, calling out the action to be done to the content, and entering new input for substitution or addition. Let's assume that re-recording the whole input over again is too tedious and not practical for some of these applications. The design problem we are looking at entails the following issues:

  1. Is the metaphor of positioning and replacing audio input appropriate for speech interface? Is it an inappropriate metaphor carried over from the cut-and-paste paradigm in a visual interface?

  2. If it is not the right metaphor, what is a better metaphor for manipulating audio content?

  3. If it is indeed an acceptable metaphor, what is an effective technique to position and select a range of audio input (or output if we count when it's played back)? What are design alternatives?

  4. Is the technique different if a small display is used with the manipulation task? (This only works if the audio input has been recognized and transcribed into text.)
Please note that the use of a dictation engine is somewhat an orthogonal issue to the design problem here. I mentioned dictation to state the design space -- that is dictation enables long text entries through speech.


CHI '97 Workshop: Speech User Interface Design Challenges