Why I Made the System Spell Names Out.

01 / FOUR WORDS WERE ENOUGH

Four words were enough

The operator did not need a conversation with the computer. Four short commands covered the work: Next, OK, End, and Back.

A smaller vocabulary made the interface faster to learn and easier to recognize in a noisy workplace. The goal was not to show how clever speech technology could be. The goal was to let the worker keep both hands on the specimens.

02 / PEOPLE DO NOT PRONOUNCE THE SAME WORD THE SAME WAY

People do not pronounce the same word the same way

In Japanese, the normal confirmation was hai—“yes.” But my own articulation is not always perfect. When the system refused a perfectly understandable answer, my reaction was immediate: why will it not recognize me?

So I taught it to accept close pronunciations such as fai and ai as the same intention. If a person’s meaning is clear, forcing a textbook pronunciation only makes the system less useful.

Recognize the person’s intention, not an idealized pronunciation.

03 / MAKE NAMES AUDIBLE, NOT MERELY CORRECT

Make names audible, not merely correct

Speech synthesis had the opposite problem. A computer could pronounce a Japanese name in a technically plausible way and still leave the listener unsure which characters it meant.

For names that could be confused, the system spelled out the critical sound. For example, it could clarify a voiced consonant by saying the equivalent of “ko with two dots” or “shi with two dots.” It was not elegant Japanese. It was unmistakable operational Japanese.

04 / THE BEST INTERFACE IS THE ONE PEOPLE CAN CORRECT

The best interface is the one people can correct

These details came from using the system ourselves and listening to the people doing thousands of repetitions. A voice interface becomes dependable only after real voices, real accents, real fatigue, and real names have had the chance to reshape it.