by Gary Wunder
From the Editor: Almost everyone agrees that Intel's involvement in bringing its resources to the blindness field is a welcome addition, potentially enhancing our easy access to the printed word. Because we want to encourage continued innovation in access technology, we have included the Intel announcement about its new technology in this issue of the Braille Monitor.
Several technology-savvy members of the National Federation of the Blind have participated in early testing of the Intel Reader, and their initial assessment is that the Intel product may fall short of the contribution we want from such technology. Gary Wunder, secretary of the National Federation of the Blind, president of our Missouri affiliate, and a technology professional in his own right, has been an early tester for the Intel Reader. In correspondence he has forwarded to Intel and HumanWare, Gary has acknowledged the potential usefulness of this device in the lives of blind people while pointing out several fundamental shortcomings that he believes must be resolved before the Intel Reader will be equipped to take its place with the best access technology on the market. We reprint below excerpts from Gary's communications with Intel and HumanWare. These reflections are by no means complete or formal reviews of the product, but they do signal concern about the ability of the Intel Reader to perform competitively at its current stage of development. The first piece was sent to officials at Intel, and the second was sent to staff at HumanWare. Some of Gary's observations in the messages are duplicative, but each has enough unique perspective for us to reprint them both. Here is what he says:
Memo to Intel
For me print-to-speech technology has opened many doors, but the most important has been to the public library and books. I'm glad to read an incoming letter, a work memo, or an agenda prepared too late to come to me in an accessible form, but books, because of their length, are not conveniently or economically handled by the human readers I employ or the pocket-size reader I carry.
My enthusiasm for the Intel Reader, which I will refer to as “reader,” was that it could shoot a book faster than I could scan it and it seemed to offer the possibility of handling everything from the smallest paperback to the largest textbook. This is not so, of course, for the typical eight-and-a-half-by-eleven-inch flatbed scanner.
The bottom line for measuring the success of any print-to-speech system is accuracy. Especially is this true for books. I cannot submit to Bookshare a book of lesser quality than I could deliver with a scanner and justify my submission by saying it took me less time to prepare. The same is true of books provided to students whether in the public schools or by disability offices in colleges and universities. The University of Missouri scans its textbooks by removing the book's binding, running the pages through a high-speed scanner and a commercial OCR package, and then rebinding or discarding the book. Improved speed over this approach cannot be achieved by the reader, and in any event students will grade the performance of the service provider based on the material and not how long it took the provider to prepare it.
Currently the reader market for the blind consists of pocket-sized machines to PC-based scanner systems. Some of us who are fortunate enough to have two systems use our scanners for books or large documents and our portable readers for on-the-go reading. For this reader to succeed in the group having two devices, it must perform as well as the scanner-PC systems or become small enough to fit in a shirt pocket and provide multiple functions as now happens with the knfbReader, which is also a cell phone providing access to voice, text messaging, and email. If the reader is excellent, it may capture the market and reduce significantly the demand for both of the systems mentioned above. For wonderful accuracy and speed, I might sacrifice the convenience of my shirt pocket. If the unit is as accurate as my scanner and PC system, I might well regain some of my physical desktop and reduce the maintenance cost I pay for all of it.
I initially looked at this machine in the same way many families look at motor vehicles. There are days when we want a fancy sports car for fun and to show off to our friends, and there are days when we want a truck to bring home the new furniture we've purchased. If we can't afford both, we settle for something in between--the family car, which may be a station wagon, a minivan, or a SUV. It will be classy enough to take into areas where trucks are not allowed and functional enough that we can use it for limited hauling. The most important function it will serve is daily getting us to and from work and taking our family where we need to go. Beyond this somewhat flawed analogy, we still get back to the question of what place the reader has in the market. If this isn't the smallest and isn't the most accurate, it may prove to be too big to displace the cell-phone reader and too slow and inaccurate to replace desktop systems.
In my week of work with it, it cannot best the accuracy of desktop systems and does no better than the knfbReader where accuracy is concerned. The knfbReader is quicker to tell me whether I captured all of the text in my shot and faster to tell me if a page is blank--a matter of tremendous significance to a blind person who may have the page wrong side up. With the Intel Reader I can shoot faster, but, in delivering the first and subsequent pages of information, it is slower than both.
It is important to differentiate between the feedback you will get from experienced users who have seen and heavily used current technology versus what you will get from someone who has never had the pleasure of a machine reading any printed document to him. To the person with no reader experience, anything which causes what appears to be a blank piece of paper to speak is miraculous. I observed this in watching testers of the knfbReader, who confidently asserted that this was the fastest, most accurate reading machine ever made for the blind. This was indeed their strongly held opinion, but it was not an informed opinion, and knowledgeable purchasers of blindness technology will in the end make up their minds about the relative merits of products based primarily on accuracy.
Given this machine's speed of image acquisition, it could be a major player if recognition were improved. If accuracy cannot be made to exceed the cell phone devices and match the scanner PC systems, it is hard for me to see a market. There is a lot of merit in the idea of a portable device for on-the-go reading and a docking station for large document reading. The question is whether this technology can marry the advantages of both by delivering a quality audio presentation from print.
I am honored to have been a part of the group chosen to test this machine, and, while I would not purchase it at this point, there are many things I admire about it, and I hope enough development occurs that it will live up to the potential which seems to exist in a quality capture station, a good processor, and all of the thought which has gone into a well-designed unit.
Memo to HumanWare
This unit is exciting because it employs the concept of a portable reader that can be made into a real workhorse when paired with the capture station. It is comfortable to hold, its buttons are very discernible, and the layout of the unit is logical and easy to understand. Its battery life is good, and its recharge time quite acceptable. Capturing images is easy using it as a handheld unit and most especially using the capture station. Regardless of how unscientific or provable the term, the machine is just cool.
All of this being true, the bottom line for me is that, in its current stage of development, this reader fills no niche in my life. My needs focus on reading the mail; reading work documents handed out too late to be scanned before meetings; and, my true joy, going to the library to get books I would not otherwise be able to read. This unit cannot replace the reading device I carry in my pocket. It is faster to shoot multiple pages but not faster to recognize or read the captured material. Although its camera and processor should make it more accurate than my knfbReader, I do not find it so. The reader cannot replace the K1000 on my desk because it is far less accurate. When it comes to books, accuracy is far more important than is the speed in acquiring an image. With the utmost respect and absolutely no sarcasm intended, I have to say that, even if this unit were given as a reward for testing, in good conscience I'd have to send it back because I don't know where I'd use it.
None of what I've said has anything to do with my hope for what the reader can become, my appreciation for the design concept of a portable unit coupled with a text-capture station, and the power and flexibility of the processor being used. Neither does it mean I am unwilling to help in ongoing testing so that the resulting unit is so fast and accurate that it is worth the extra size to carry around, and its speed of image acquisition and accuracy let me free up some of my physical desktop and avoid the yearly maintenance cost of updating other print-to-speech programs.
I see all kinds of potential, but most of it at this point is theoretical. This unit simply is not comparable to other commercially available OCR products both in and outside the blindness field, and I hope Intel and HumanWare realize that blind people with existing technology will not replace or supplement what they have with this unit. Further, blind people who have not yet enjoyed the blessing of turning print into the spoken word deserve much better recognition than you can offer here.
Intel must first bring better OCR to the product. When taking a shot, blind people should know rather quickly whether the page appears to be blank, whether it appears text is cut off, and what part of the page is missing. If possible, the engine should provide an assessment of the confidence level the engine has assigned to the page. The latter suggestion may be beyond the capability of the processor, but certainly the first two are not. Optionally the orientation of the page should also be announced because sometimes I'm trying to arrange papers for the review of others, and they demand more from me than sheets that are upside down.
It would be desirable if the unit could shoot pages while reading previously processed material. Stated differently, I'd like to be listening to a book while I continue to photograph it. If processor speed will not allow for this parallel activity, the unit should at least provide the feedback about each page listed above.
As important as these features are to the unit, the central issue is still accuracy, and, unless we have it, all the bells and whistles I might suggest still won't sell or serve the competitive blind person who wants and needs to read print. The last sentence is the most important one I wish to convey in this note. I just talked with a staff member at Intel who has looked at my images and resulting text and says she believes the reader is being used properly and the quality is simply what the unit, in conjunction with the capture station, can now produce.
Last, and certainly most subjective, is the issue of the text-to-speech being used. My perspective is that there are some very responsive engines such as Eloquence, DoubleTalk, and Keynote Gold which no one would mistake for human speech but which, to the trained ear, become quite comfortable and provide rapid and very understandable output. At the other end of the spectrum are some very human sounding text-to-speech programs which aren't nearly as fast or responsive but are preferable for the reading of material intended primarily for enjoyment. My take on the Intel Reader is that its speech falls somewhere in the middle, but this is not a compromise which serves anyone. The current speech software has the disadvantage that it does not really sound human, nor is it rapid or responsive. It represents an attempt to provide more than the robotic speech of the synthetic speech most of us use on personal computers but falls far short of speech offerings from ScanSoft, AT&T, Neo-Speech, or L and H. I understand the appeal in marketing for a system which does not sound like a machine, but I think having an option which is fast and responsive will attract serious users of text-to-speech technology. Continuing to search for something more human-sounding will also be helpful if the processor will allow for it; it will add not only to initial appeal but to reading for enjoyment which this unit may eventually provide.
I know I could have returned my machine without a response, but I'd like to see this machine mature and earn a place in the homes and workplaces of blind people. Our reading needs are not so well met that there is no place for another unit, but any new offering should at least meet and preferably exceed what is currently available. If the unit could do what it has the potential to do, the reader is something I'd be a fool to pass up. I believe HumanWare and Intel can make this machine become something really useful to blind people, and I stand ready to help with future testing if this is desired.