Originally presented at the AVIOS 2000 conference, May 23, 2000, in San Jose, California. This paper is found as pp. 9-14 in the proceedings under the title "Internationalizing Speech Applications."

LOCALIZATION OF SPEECH RECOGNITION SOFTWARE FOR JAPAN

Over-the-phone speech applications are less easy to translate to other languages than web applications or traditional desktop GUI applications. Using a mouse and keyboard requires training, but using the phone and speech itself is a familiar experience for which callers already have many linguistic and cultural expectations. Speech applications must adapt to these expectations, since longstanding habit cannot be trained away, even for the most cooperative callers. The adaptations required for internationalizing speech applications often go far beyond substituting one language's vocabulary for another.

Examples from the translation of two applications from English to Japanese show that even apparently simple, common items—names, numbers, answers to simple questions—pose a wide range of surprising challenges for both speech input and output. However, addressing the problems raised by one language can lead to a better solution for speakers of all languages. Internationalizing an application also reveals some of the peculiarities of English and American English, shedding light on the voice user interface design process.

[1] Apple Computer, Inc., Inside Macintosh: Volume VI. 1991, pp. 2:4-12.

Introduction

Application developers are usually cautioned to consider international users and markets early in the design process. In the domain of graphical user interfaces, internationalization means translating labels, allowing right-to-left text and extra room for languages that are not as terse as English, dealing with new currencies and date formats, and checking icons for cultural appropriateness. [1] Rarely if ever is the logic of the application affected or the look and feel significantly altered.

Why is this not true for speech applications? Many common items in a speech recognition application—dates, times, names, money amounts, yes/no answers—exhibit some of the least logical traits of human language. Take dates as an example. Nearly every culture and language provides a completely numerical method for writing dates. Hardly anyone uses these numeric forms when speaking, however. Instead of saying "five, thirty, two thousand," most people would say "Tuesday next week," and the people or programs who need the numbers can figure them out for themselves. This does not make things easy or uniform to program, but it does make for an interesting challenge for the developer who wants an application that sounds and behaves naturally across cultures.

The examples here are based on two demonstrations developed for a multinational company. Both applications were developed originally for American English and then localized for Japan.

The first system gives order status reports and deals with numbers, money amounts, dates, and company and product names. It was not originally designed to be translated.

The second is a dialer, or auto-attendant. The dialer was designed to be localized for Germany as well, but only the English and Japanese versions were completed. The dialer handles personal names, department names, phone numbers, and commands. It provides separate phone numbers for internal and external callers, and web-based interfaces in English and Japanese. Unlike their speech counterparts, the web interfaces were localized simply by replacing all user-visible text strings. Although each language has a separate phone number, all languages share all data.

The most interesting problems arise with names, numbers, dates, and phone numbers. Commands and proper names had straightforward translations or transliterations. To translate these applications, I worked with a translator who is a native speaker of Japanese, familiar with both US and Japanese business practices. I had completed a class in third-semester Japanese myself. This prepared me to ask questions, but it is difficult to imagine how a speech application could be successfully localized without a native speaker who is not only fluent in the language, but observant of cultural differences and knowledgeable about business practices.

I will review cultural differences, and look in detail at personal names, numbers, and phone numbers, before offering recommendations about designing and building speech applications for an international audience.

Localization vs. Translation

Localization is not translation. A statement appropriate for one culture translated directly to another may be too abrupt, too patronizing, or considered redundant in another. Some concepts may not translate at all. The dialer is a case in point. A Japanese company that started using a machine to answer its calls would alarm its customers and others, because it would seem to be indicating that it is going through such severe financial difficulty that it can no longer afford a receptionist. The internally accessible version, however, was appealing to employees in Japan because they wanted to have calls routed without bothering a receptionist. German colleagues also pointed out that the German office would be very unlikely to use a publicly available auto-attendant for the same reasons.

There are also differing expectations about how machines should speak. As an example, speech applications for American audiences take broadcast media as a model for how they should sound: friendly, but brief and to the point, using as few words as possible. Though talking cars and vending machines have been brought to market in the US, Americans have rejected these devices as mechanical and impersonal. By contrast, Japan is full of talking machines and other automated announcements—the automated reporting of train arrivals is considered useful and helpful information. There are even shops in Japan that use motion detectors to trigger recorded shouts of irasshaimase (welcome). Invariably, because they are providing customer service, these recorded announcements use polite language, even when shorter, more efficient alternatives are available. Often, the useful information is surrounded by words of thanks and warnings to be careful. The background noise of polite recordings is essential to any Japanese train station. Female voices dominate, because this is considered more polite. The English dialer uses a male voice, but the Japanese dialer unquestionably needed a female voice.

When speaking to machines, Japanese callers are likely to be much less wordy in their responses, and the filler words are more uniform. An American asked for a quantity might say "forty-five", "I'd like forty-five," "forty-five of them," or "forty-five, please." The translator and callers offered few alternatives to the equivalent of "forty-five units." In Japanese, sentence subjects are frequently omitted, no verb is really necessary in this case, and the equivalent of please would be incorporated into the verb ending. There is nothing to add to this simple statement. This makes some of the Japanese recognition grammars much simpler than their English counterparts. Below in the section about numbers, I explain why Japanese callers would say "forty-five units" instead of just "forty-five."

Names

Name order

The Japanese say their family name before their given name. (To avoid confusion, I'll use the term "family name" for what English speakers would call a last name or surname, and "given name" for first name.) Even people who have Japanese business associates may be unaware of this, though, because when speaking English, the Japanese adjust to Western ways.

You might think, therefore, that swapping the family name and given name would be all you have to do. This would be true, except that in Japanese practice, a Western name is still given in its Western order. How do you tell if someone is Japanese? The database included an employee's location, but this field was not sufficient for distinguishing Japanese from Western names—the Tokyo office had employees from all over the world, and there were many Japanese and Japanese-Americans elsewhere. Japanese-Americans pose yet another problem: some Japanese callers will say an obviously Japanese name in Japanese order, but others, knowing that the person doesn't even speak Japanese, will use the Western order.

The solution to this problem was to check to see if the name followed the rules of Japanese, and to put them in Japanese order for the Japanese version. To accommodate Japanese-Americans, Western order was also allowed for Japanese names. So Western order was allowed for all names, and Japanese order was also allowed for names that sound Japanese (technically, they just followed the rules of standard romanization).