I ARRIVE punctually at the university lab in Tokyo to find the subject of my interview, a certain Waseda Talker-san, waiting for me in the far corner. I鈥檓 immediately struck by the petite face and cute nose, rather spoiled by a greyish, pockmarked complexion. Even more unsettling are the lifeless eyes that have me wondering what鈥檚 going on behind them. Not much, judging from the flat, unchanging expression.
Waseda Talker is a robot, albeit a rather special one. As the name suggests, this robot can speak. No big deal, you might say, given that electronically synthesised speech has been around to irritate us for decades. But in this instance, researchers at Waseda University, which has a distinguished record in robotics stretching back 30 years, have shunned the synthesised approach to android-speak. Instead they have created a machine that can mimic human speech under its own steam 鈥� or, in this case, its own compressed air.
When graduate student Kotaro Fukui clicks an icon displayed on the computer, WT-5 (the fifth build of Waseda Talker) stirs into action. Motor-driven diaphragms force air from twin plastic tanks up through an artificial vocal tract, over the tongue and out of the mouth and nose. Simultaneously WT-5鈥檚 mouth flexes into life as its lips open, close, stretch and protrude, all while articulating the five Japanese vowels in sequence 鈥� 鈥渁/i/u/e/o鈥� 鈥� in a voice midway in timbre between a man and a voice synthesiser. Fukui clicks another icon and the robot chants out 鈥渄a/di/du/de/do鈥�. It reminds me of an old primary-school teacher laying down the law during reading lessons.
Advertisement
The machine鈥檚 creator, Masaaki Honda, is a computer scientist at Waseda whose interest in biomechanics has led him to fix his sights on building an autonomous talking robot. Why? He wants to understand better what goes on in our heads when we speak. Such insights might be used to help people with speech handicaps and to create tools for speech training and language learning.
While that goal is some way off, the robot has speech researchers intrigued. 鈥淟ooks interesting to me,鈥� says Tecumseh Fitch, a specialist in vocal communication at the University of St Andrews in the UK. 鈥淚鈥檇 like to see them make a talking dog or monkey with this technology. A talking Aibo.鈥�
But WT-5 is more than just a toy: it is part of a trend to build robots that communicate and interact with people more naturally. In this it competes with machines such as Hideyuki Sawada鈥檚 talking and singing robot at Kagawa University in Japan (New 杏吧原创, 4 May 2002, p 24). And while babbling bots may seem like an annoying prospect, there are other applications. Honda thinks the research could yield cellphones that represent the vocal movements of a speaker and transmit this data to a voice synthesiser on the other end, thereby reducing the bandwidth needed to communicate. The projects might also lead to better ways to control artificial vocal cords for use by people who cannot now speak.
What Honda and others seek to discover is how the human brain controls our speech articulators 鈥� the lips, tongue, vocal cords and such 鈥� when we talk. They know that when we get an idea we want to express, neural commands summon signals in the motor cortex, the part of the brain that controls our voluntary muscles. 鈥淲hat we don鈥檛 clearly know is the sequencing of all this, or how the different circuits in the brain work together to produce speech sounds,鈥� says Kiyoshi Honda, head of biophysical imaging at ATR Human Information Science Laboratories near Kyoto, Japan. 鈥淎nd we won鈥檛 understand it exactly until we can reconstruct the brain circuitry and machinery of speech.鈥�
Speaking out
To grasp the complexity of this challenge, consider what happens when we speak. First, our lungs push air up the windpipe and past the vocal cords. When the cords are tensed, the airflow causes them to vibrate and produce sound. This is how 鈥渧oiced鈥� sounds such as 鈥渄鈥�, 鈥渂鈥� and 鈥渧鈥� are made. When we utter unvoiced sounds such as 鈥渢鈥�, 鈥減鈥� and 鈥渇鈥�, or when we whisper, the cords relax and we simply push air out through the mouth. In both cases we manipulate the shape of our mouth, including the lips and tongue, to produce the range of sounds that make up spoken words.
That鈥檚 where the robot comes in. 鈥淭he Waseda Talker researchers have built a mechanical system to simulate physiological processes of speech production, which has never been attempted before,鈥� says Kiyoshi Honda. Though there are already models that suggest the vocal tract requires certain patterns of neural commands to generate speech sounds, he hopes the Waseda project will yield its own theories.
It all began in 1998, when Masaaki Honda and a group of engineers, clinicians and acousticians decided to build a mechanical speech synthesiser. Using MRI images of human speech organs as a guide, the group constructed a talking head equipped with artificial vocal cords, tongue, teeth, lips and a nasal cavity. Each of the articulators was capable of various degrees of movement. The tongue, which like the lips was made of soft synthetic rubber, was the most flexible. Electric motors and crank mechanisms operated levers, springs and wires attached to the articulators to provide the movement.
The group adjusted the movements of each articulator until the robot could produce recognisable vowel sounds 鈥� though Honda would be the first to admit that its voice was hardly natural. One reason is that while its lips could be stretched and made to move up and down, they were unable to protrude, something our lips do when we make the 鈥渙o鈥� and 鈥渨鈥� sounds. The vocal cords were made from rubber plates with a passage for air to flow through. A motor pulled or pushed on the cords to stretch or loosen them, producing voiced and unvoiced sounds in turn.
The design was too simple to match the richness of human sounds, but over years the researchers improved the robot鈥檚 articulation by redesigning its palate, tongue and lips to be more flexible. A new way to control the vocal cords was introduced, using a second motor in the stretching and loosening process. Such designs enabled the bot to articulate vowels more naturally and to add sounds such as 鈥渟鈥� and 鈥渕鈥� to its repertoire. By 2004, the robot had produced all 50 Japanese speech sounds or 鈥減honemes鈥�.
The group also devised a system that enables the robot to mimic a few words, such as hassei (Japanese for 鈥渟peech鈥�). Hassei contains a pause between the two syllables, and the first syllable is voiced, while the second is unvoiced. The robot learns by copying a human, though it needs help along the way. Sound-analysis software breaks the word spoken by the human into its acoustic components, such as pitch, volume and tonal characteristics. These are used as guides for the robot鈥檚 utterance, which the analysis software compares with the original. The researchers then fine-tune the speech by adjusting the controls of the lips, tongue and vocal cords. After many adjustments, what results is fairly close to the human sound, and a record of the settings is stored, ready to be called up again.
It鈥檚 a painstaking process that cries out for automation. So the researchers are developing a computer model for voicing phonemes that could enable the robot to mimic new words all by itself. The research has not progressed past being an engineering exercise, says Minoru Asada, a robotics expert at Osaka University in Japan. 鈥淲hat I expect them to do in the future is attack the issue from the viewpoint of a baby that gradually learns to speak.鈥�
That may be a decade away, and the robot has yet to overturn any theories of speech production. But as I contemplate WT-5 sitting silently in the corner, part of me is relieved that babbling bots are not yet reality. Let鈥檚 enjoy the peace and quiet while we can.
