Voice Biometrics Conference Convenes in NY
How topical. The annual Voice Biometrics Conference begins today in New York. This year's theme: Benchmarks, Use Cases and Real World Experience.
Registration is $699., but maybe they have media passes available. One interesting talk may be that of Alexey Khitrov, of the Speech Technology Center which is a "gold" sponsor of the conference. (Its program is SpeechPro):
Reality Check #1: Lessons Learned from Forensics and Law Enforcement
What other programs do law enforcement use for speaker recognition? [More...]
Here's an interesting 2009 declassified FBI report on voice biometrics. The report notes investigatory voice biometrics may not be ready for courts --at least federal ones that use Daubert as the standard for admission of expert testimony:
The first agreement was on the meaning of the phrase - “investigatory voice biometrics”. The first word of this phrase, “investigatory”, was taken to encompass the use of speaker recognition technology in criminal and intelligence investigations and analysis.
In US Federal Courts, the admissibility of scientific evidence is determined by the presiding judge, who is guided by Federal Rules of Evidence (FRE) 104  and 702 , among others. FRE 702 notes the Daubert Criteria, which states the following factors must be met: the technique has been tested and subjected to peer review and publication; has a known error rate and standards controlling its use; is generally accepted in the scientific community .
The Symposium committee members believe that automatic speaker recognition technology has not yet reached the maturity to satisfy the Daubert Criteria...Future research and evaluation are needed to advance speaker recognition technology to satisfy scientific-evidence admissibility requirements for the US Federal Courts
In addition, there is a common consensus about the uncertainty of the error rates of the speaker recognition performance when operating under variations due to the speaker (e.g., stress) and variations not due to the speaker (e.g., channel distortion).
On p. 14 (its numbering) the report addresses the process for comparing a 9/11 call to a sample of a known person:
4.2 FBI Scenario 1: One-to-one Comparison
The Forensic Audio, Video, and Image Analysis Unit (FAVIAU), Operational Technology Division of the FBI, will receive a request from one of the field offices for a voice comparison examination to be conducted between an unknown voice on a 911 call and the known voice of a subject on a recording obtained by law enforcement officers.
...Currently automated comparison is performed in criminal cases solely for research purposes and only if the evidence is deemed sufficient for such an examination.
Another symposium to address the current state of voice biometrics was held in 2010. The declassified report is here, but its very technical.
Here is the latest draft (March 12, 2012) of the Investigatory Voice Biometrics Committee Report from the National Institute of Standards and Technology (NIST). On page 50-52 (its numbering) the draft addresses intelligibility, emotional state and impairment, and speaker styles (crying, chanting, etc.) and "vocal effort."
Here's the 2011 report, National Biometrics Challenge from the Office of Science and Technology Policy. It says on page 14 (image provided as it doesn't allow cut and paste):
Does law enforcement use EasyVoice Biometrics? According to its website, the software was just released and didn't start shipping until March, 2012. (It costs $5,000.) I was wondering how Tom Owen (whom I have used as an expert in the past and found to be quite professional) got an advance copy since he testified at a court trial in January that he had used the program the year before. Turns out, according to Whosis, he is the administrative contact for the EasyVoice Biometrics website. If he has a proprietary or financial interest in the program, or helped develop it (which I don't know that he does or did) should he have disclosed it to the Sun Sentinel?
As for the Connecticut state court's admission of his biometric voice analysis at the prior murder trial, it should be noted that the defendant in that case was pro se, without a lawyer, and likely didn't know how to effectively challenge it. Also, it's unclear how much the analysis factored into the jury's verdict since there was DNA evidence incriminating the defendant. Her DNA was found in a blood stain on a bathroom sink handle in the victim's condo.
Jurors on Friday morning asked the court to replay testimony from Michael Bourke, a supervising DNA analyst in the state's forensic crime lab. Bourke testified the estimated frequency of someone other than Davalloo leaving that DNA profile on the sink handle was 1 in 8.5 million.
Also, the 911 case in Davaloo did not involve comparing a scream to spoken words, although I'm not sure it that matters.
On the 911 recordings, a woman tells a call-taker that a man attacked her neighbor on Harbor View Drive. The call was made from a pay phone at a fast-foot restaurant on Shippan Avenue.
The second expert hired by the media to opine on the screaming in the background of the 911 call in the Trayvon Martin-George Zimmerman shooting investigation pooh-poohed biometrics.
Ed Primeau, a Michigan-based audio engineer and forensics expert, is not a believer in the technology's use in courtroom settings. He relies instead on audio enhancement and human analysis based on forensic experience.
Hopefully whatever expert the state uses will come up with a better explanation than Primeau did for saying the voice was not Zimmerman's:
"I believe that's Trayvon Martin in the background, without a doubt," Primeau says, stressing that the tone of the voice is a giveaway. "That's a young man screaming."
On his own website, Primeau cautions that his opinion is not for legal purposes.
As an audio forensic expert I want to make clear that this is not formal voice identification and not meant for legal purposes.
He also may not have done a forensic comparison at all. He says:
If I were to conduct a formal voice identification test, here are the steps I would take:
....4. Ask for a speech exemplar from George Zimmerman screaming for help
5. Conduct a voice identification test.
So unlike Mr. Owen, Primeau thinks a test should compare a scream to a scream. Since he doesn't have an exemplar of Zimmerman screaming to compare to the 911 call with the scream in the background, he can't do a test the way he thinks it should be done.
It will be interesting to see whether the State's Attorney in Florida retains experts who use a biometric program to analyze the scream in the Trayvon Martin case, or goes with more traditional methods of speaker recognition. And whether the experts it does retain are able to draw any conclusion at all as to who screamed.
In the meantime, maybe the case will be brought up at the biometrics conference tomorrow and there will be some tweeting about it.
|< The Demise of The Black Caddy | Wankers >|