By David F. Salisbury
Oct. 23, 2000

Thirty years ago, Stephen Silverman was a clinical psychologist on the staff of the Brookdale Hospital Medical Center in New York. As he interviewed patients in the psychiatric emergency room, Silverman found that the sound of some voices literally caused the hair on the back of his neck to rise. As he analyzed the voices that created this sense of alarm, he concluded that it was caused by a distinctive quality in the pattern and tone of the voices of individuals likely to attempt suicide in the near future.

Silverman When he went into private practice in Connecticut, he became well known for his work with suicidal and psychotic patients. His wife, Marilyn, is also a psychologist in private practice and has dealt with some suicidal patients as well. Their experience seemed to confirm the validity of his earlier observations. So, in the 1980s, the couple decided to investigate the phenomenon on their own.

"I felt we had to answer three questions: What was it that I heard? If I heard it, is it real? And if it's real, is it measurable?" Stephen Silverman said.

After years of study and research, the psychologists finally have objective evidence that what he hears is real and considerable encouragement that it may be measurable as well. An interdisciplinary research team at Vanderbilt University working at the Silvermans' behest, has identified several acoustic features in the voices of people recorded shortly before they either took their own lives, or tried to do so without success, that differentiate them from recorded voices of normal individuals and those who were from severe depression, but were not suicidal.

"The research done at Vanderbilt supports the hypothesis that near-term, suicidal voices have distinct characteristics," said Marilyn Silverman. "And they support the proposal that this investigation should be continued with even greater scope and intensity."

Suicide is currently the eighth leading cause of death in the United States and the third leading cause of death among 15-24 year olds. More than 30,000 people kill themselves every year and 500,000 Americans make attempts or threats that require evaluations in hospital emergency rooms and/or hospitalization. Moreover, suicide rates are on the rise.

One of the biggest problems in suicide prevention is determining the degree of suicide risk of individual patients. Current assessment methods use personal histories, clinical examination and psychological testing to judge a person's risk. Because this kind of information takes a long time to gather, it is not sufficient for urgent situations that require immediate judgments. So a rapid method to assess suicidal intent by analyzing the qualities of voice patterns systematically could be a powerful new tool in the fight to prevent suicidal deaths, the Silvermans maintain.

"The work is promising and has potential clinical value," said Jo-Anne Bachorowski, assistant professor of psychology at Vanderbilt, who has been involved in the project. However, she cautioned that a great deal more research must be done before this promise can be realized. "It's crucial to recognize that the relationships between tone of voice and any psychological state or condition are extremely complex. So it is very important to test possible acoustic markers of suicidality with a wide variety of comparison groups and prospective research designs."

In the early 1980s, the Silvermans began to collect and analyze recorded suicide notes and interviews made shortly before suicide attempts. Stephen Silverman enlisted the help of a musician friend to begin identifying and categorizing the qualities that these voices had in common. They came up with descriptive terms such as dull, lifeless, metallic, hollow, lacking the sounding board represented by the body. They became convinced that the voice itself -- independent of what a person was saying -- contained important information about a person's immediate psychological state, including the decision to commit suicide. Friend and collaborator Ray Terlaga of Litchfield, Conn., performed some preliminary acoustic analyses and the Silvermans applied for and received three conceptual patents on the idea.

Vanderbilt researchers get involved

In 1994, the Silvermans visited Vanderbilt and described their project to professors Thomas Harris, chair of the biomedical engineering department, and Richard Shiavi, an expert in signal processing. The Vanderbilt researchers arranged adjunct appointments for the Silvermans, who are the clinical faculty at Yale University School of Medicine. Professor Shiavi assembled a team of experts, including Bachorowski, Mitchell Wilkes, associate professor of electrical engineering, and Ralph Ohde, professor of hearing and speech, to work on this unusual problem.

Early analysis was performed by doctoral student Dan France, who is currently a consultant at the Center for Clinical Improvement at Vanderbilt. One of his first challenges was figuring out how to deal with the wide variation in quality of the recordings, which were made "in the field" using tape recorders of varying quality, without the precise control of recordings made in the laboratory. He analyzed the different recordings and used a variety of techniques to adjust them so that they were fairly uniform in quality.

Once that was done, France analyzed the characteristics of the speech patterns of 34 normal subjects, 59 suffering from depression and 22 that were high-risk suicidal. He performed a "macro" analysis, looking at an array of speech characteristics averaged over 2.5-minute periods from each subject. He found several features that appear to differ in the voices of normal, depressed and suicidal individuals.

One of the features involves the dominant frequencies of the voice, called formants, which are determined by the physical characteristics of the vocal tract. Another involves the power spectral density of the voice, a measure that quantifies the amount of energy in different frequency ranges. France found that certain changes in these two features, taken together, correctly differentiated between depressed and suicidal individuals 80 percent of the time. The results of his analysis were published in the July 20 issue of the Transactions on Biomedical Engineering published by the Institute of Electrical and Electronics Engineers, Inc.

"It's amazing that we can get such a high level of accuracy, particular given the large variety in the quality of the recordings that we have to work with," said Shiavi.

More recently, Asli Ozdas, a Vanderbilt doctoral student, has been investigating the second-by-second variations in the rate of vibration of the vocal chords, called the fundamental frequency. In 10 normal individuals, she found that changes in the fundamental frequency occur relatively smoothly at this time scale. But in 10 suicidal individuals, she has found a significant increase in the extent to which the fundamental frequency jitters or jumps around erratically. The difference in jitter is pronounced enough to allow her to correctly differentiate suicidal from normal subjects 80 percent of the time, Ozdas reported at the IEEE Conference on Systems, Man and Cybernetics in Nashville Oct. 11.

"The jitter result is very interesting and exciting because it is unexpected," said Marilyn Silverman. "We don't know how increased jitter may contribute to what we hear in the voices."

Despite such an apparently high detection level, considerably more research is needed to determine whether acoustic characteristics of this sort can be used as a reliable means for detecting individuals at high risk of suicide, the researchers agree.

One serious limitation of the current study is the small number of cases that have been analyzed. Tapes of suicidal individuals are not easy to find. The couple hopes to locate additional recordings from near-term suicides for future studies. However, they argue that the intense study of even a small number of "robust" samples has an important role to play in this kind of research effort.

A key problem with any practical detection method is ensuring that the number of "false positives" and "false negatives" are acceptably low. In this case, a false positive is when the detector mistakenly identifies a subject as a high-risk suicide. A false negative is where it fails to identify a high-risk suicide. So, before it can be used for a detection system, the researchers must determine whether consistent acoustic characteristics accompany suicidal impulses that arise from different preconditions, such as depression, anger or drugs. They must also ensure that other mental states do not have the same effects on voice patterns as the decision to commit suicide.

Until now, the research has been funded at a relatively low level by private donations that the Silvermans have collected, and by Vanderbilt University.

"We are talking about human tragedy," said Stephen Silverman. "The suicide attempt rate at universities is very high. It's becoming an epidemic. We will feel we have done our part if, down the road, we can help save lives by developing a rapid, non-invasive way to assess suicide risk that can be used not only in hospital emergency rooms, but also in student health centers or hotlines where few other resources for assessment may be available."