Stupid Question ™
Jan. 25, 2001
By John Ruch
Q: How can voiceprint machines identify a voice even when it’s disguised?
A: Actually, there’s no evidence that they can. Indeed, even proponents have dropped the term “voiceprint,” because it incorrectly suggests the subjective practice is as reliable as fingerprinting.
“Spectrographic voice analysis,” as it’s called today, was first proposed during World War II. But it came to fruition only in the 1960s, when Bell Labs physicist Lawrence Kersta set out to ID bomb-threat callers for the New York Police Department.
Using a device called a sound spectrograph, he claimed to ID voices with 99.65 percent accuracy. Kersta trademarked the word “voiceprint,” founded Voiceprint Laboratories and established an International Association of Voice Identification. Like the fellow quasi-science of polygraphy, voiceprinting relies on arbitrary judgments and an unproven, reductionist hypothesis: in this case, that every human voice is unique and can be reliably discerned by a machine.
Voiceprinting is admissible evidence in only a handful of states, and then mostly as a tool of corroboration rather than primary identification.
Since the voiceprinting method is strong biased toward giving a positive result, this helps explain its claims of high success rates: in most cases, it’s used to identify a single, already known suspect, who in the US is statistically likely to be found guilty simply by going to trial.
The method begins by getting a recording of a suspect’s voice to compare to a source recording. Whenever possible, the suspect is made to say the exact same words in the exact same way—even with the same disguise, if applicable.
First the examiner listens to the recordings, noting similarities (not differences; it’s biased toward positive ID) and selecting especially close ones to run through the spectrograph. This device records sound waves as a printout that shows the frequency and intensity of the sound. The examiner compares printouts of the sample sounds for similarities and differences in the squiggly lines; only at this point do differences count.
The results of this subjective process are then rated on a subjective scale: twenty similarities and no differences is a positive ID; fewer than 20 and no “unexplained” differences is a probable ID. The error rate of voiceprinting is unknown, though scanty studies suggest anything from 6 to 29 percent.
It probably helps that most cases brought to examiners (about 60 percent) are rejected for having “poor” recordings or “not enough information.” In the similar, and better investigated, world of computer voice-recognition programs, difficulties are rampant and manufacturers are happy with an “equal error rate” in which the number of false positives equals the false negatives.
Since the human voice has a large behavioral element, there is probably no error-free way to ID it. And almost no research has been done on the effects of illness, emotional states and deliberate disguise on the method. Indeed, in 1974, California tossed voiceprinting out of its courts specifically because there was no evidence that it could ID disguised or mimicked voices.