Voice forensics experts cast doubt on Orlando Sentinel analysis of Trayvon Martin 911 tape
Voice recognition experts who spoke to The Daily Caller questioned the methodology and conclusions of a voice-identification analysis published by the Orlando Sentinel on March 31.
Using two different forensic methods, the Sentinel’s chosen authorities determined that a scream heard in the background of a widely aired 911 call was not that of George Zimmerman, the man who fatally shot 17-year-old Trayvon Martin on Feb. 26
A police recording of the call captured part of the struggle that ended Martin’s life.
The Sentinel’s reporting has landed a blow against Zimmerman’s self-defense claims, suggesting that it was Martin, not Zimmerman, who shouted for help before the fatal shot struck the teenager in the chest.
Under contract by the Sentinel, Owen Forensic Services founder Tom Owen compared Zimmerman’s voice on his initial 911 call with the scream heard in the background of the later call. Using his own proprietary biometrics software, Owen determined that the two voices were only “a 48 percent match.”
Owen said a positive identification would involve a number over 90 percent. “[Y]ou can say with reasonable scientific certainty that it’s not Zimmerman,” he told the Sentinel. (RELATED: Full coverage of the Trayvon Martin shooting)
The scream is central to the case. If Zimmerman was the one screaming, his statement to police that he shot Martin in self-defense after being attacked would be more supportable. According to eyewitnesses and police reports, after Zimmerman shot Martin he said no one responded to his pleas for help. One unidentified eyewitness told police that he saw Martin on top of Zimmerman, and that Zimmerman was yelling for help.
But other media reports suggested that at least one eyewitness believed Martin cried out for assistance before he died.
Comparing apples to oranges
Dr. James Wayman, a San Jose State University expert in the field of speech science, told The Daily Caller that he questions the grounds on which Owen based his analysis.
Wayman also said he would be willing to testify against the admissibility of Owen’s findings on the grounds that they don’t meet the criteria required for evidence in federal courts.
“There is no history of, or data on, the comparison of a questioned scream to a known speech sample,” Wayman said.
The problem, he said, is that the two voice samples were recorded in difficult acoustic conditions over different cell phones.
“Even if we were to have Mr. Zimmerman recreate the scream under identical conditions with the same cell phone,” Wayman explained, “it would be difficult to attribute the scream to him without a sample of a similar scream from Mr. Martin under the same conditions. This is clearly not possible.”
Reached for comment, Owen told TheDC that he has conducted his own study — “The Owen Study” — of more than 400 different pitches, screams, and voice disguises. The study is unpublished.
He explained that he has attempted, without success, to obtain a “voice exemplar” from Zimmerman, consisting of recordings of both his speaking voice and a scream.
And Wayman, he said, “assumes that the voice software is not able to make a determination on each voice independently.”
Wayman fired back in a later email exchange. “There is no accepted standard regarding metrics for voice comparisons,” he insisted, “either if done forensically or using automated comparison software.”
‘Naïve’ voice recognition
The Sentinel also contracted with Ed Primeau, a trained audio engineer and registered investigator whose expert testimony has been used in dozens of criminal court proceedings. Primeau used a more intuitive approach to determine that Zimmerman was not the person heard screaming on the 911 call.
“That’s a young man screaming,” Primeau told the Sentinel.
Comparing the human voice to a symphony full of varying timbres, Primeau wrote on his blog that the “male voice yelling for help … cracks like teen male’s does when going through puberty.”
Dr. Philip Rose of the Australian National University told TheDC that scientific experts refer to Primeau’s method as “naïve voice recognition.” His influential 2002 book Forensic Speaker Identification draws a major distinction between naïve and “technical forensics” voice recognition.
“Naïve voice recognition is so prone to error that it is acknowledged that it is worthless as evidence,” Rose said via email.
A forensic expert’s job, he said, is to assess the strength of evidence, not to estimate the probability of a hypothesis. And “the value of the evidence depends … on the similarity of the samples.”
In a properly conducted analysis, he told TheDC, “you would still have to do the comparison using screamed and phone samples, with many speakers.”
One voice authentication expert whose work is commercial in nature told TheDC that screaming, stress, and a recording’s audio quality can “wreak havoc” on voice biometric software and its ability to interpret data.
And speaking of Owen’s findings, another industry insider said that “a legitimate biometrics expert would likely refute the contentions” and suggests that these were “incendiary publicity plays.”