"You do not have to say anything, but it may harm your defence if you do not mention, when questioned, something you later rely on in court. Anything you do say may be given in evidence." We have all heard this 1,000 times yet we barely give a thought as to what may happen to all the recordings that the police make of their interviews. Or indeed to the somewhat more mundane equivalent: "This call may be recorded for training purposes."
However, without your permission – or even your knowledge – your recorded voice may be about to play a key role in the race to fingerprint the human voice.
Fuelled by 9/11, spurred on by the advance of our digital society and made possible by raw computing power, the development of increasingly sophisticated automated speaker recognition systems (ASRS) are now bringing the prospect of a "voiceprint" enticingly close, threatening to make the skilled voice scientist redundant. These automated systems, already widely used by police and intelligence services on the Continent, can in as little as 15 minutes use a background population of voices to make a statistical judgement on the significance of any similarity or difference between the voice of the criminal and that of a suspect that could have taken a human 15 hours to complete.
"September 11 was the trigger for this as, after the attacks, the police and intelligence services realised that while there were so many recordings of the voices of the terrorists they didn't have the technology they needed to extract information from them," Antonio Moreno says. Moreno is the technical director of Agnitio Corp, which was spun out of the Technical University of Madrid in 2004 and provides forensic automated speaker recognition systems, such as its market-leading Batvox, to the police forces of more than 20 countries, including Germany and the US but not yet the UK.
"By the time of Spain's own 9/11 [the Madrid train bombings of 11 March 2004], Batvox could be used to identify some of the men behind the bombings as, although they wore masks on YouTube, they spoke naturally."
For Professor Peter French, founder of the UK's leading and oldest forensic speech laboratory, JP French Associates, the bugging, recording and identification of people traffickers, drug dealers and terrorists was only the beginning of this revolution.
"The ubiquity of mobile phones means that almost the first thing you do if you are attacked is call 999, and as all 999 calls are recorded a lot of people inadvertently record their rape or mugging and capture their attackers' voices," French says.
Now, though, the "great quest" is to fingerprint the human voice and "many engineers keep telling me that all they need is more time to tweak the algorithms and they can achieve full accuracy", French says. Francis Nolan, Professor of phonetics in the Department of Theoretical and Applied Linguistics at the University of Cambridge, agrees that the balance is shifting towards these automated systems due to the technical advances that have made them possible. The importance of speaker identification has grown for the simple reason that "it's not the amount or nature of crime that has changed, it's just the sheer amount of recorded material that is now available".
Nolan adds that "while on the Continent the police are more likely to use an automated system, in the UK the tradition has been to use a skilled dialectician", who would analyse one at a time the sound of the vowels and the even the rise and fall of the voice, its melody, through a complicated system of notation called the International Phonetic Alphabet.
Later, acoustic tests were introduced that allowed the dialectologist to measure the different elements of the speech signal and so extract information that was beyond the ability of humans to hear. Even with the help of technology to run these tests, the role of the specialist remains largely the same: to make a judgement as to whether – for example – any differences between two recorded voices were down to smoking, drugs, flu or even whom they were talking to, or whether they were the voices of two entirely different people.
Now, Nolan says, "we are beginning to augment the human element still further through the introduction of automated systems such as Batvox", which by analysing the speech signal analyses the characteristics of each human's vocal tract and comes up with a statistical model that can compare an unknown voice against voices coming from known speakers regardless of what they are saying. Batvox, for example, then produces a likelihood ratio, much like a DNA profile does, to suggest how significant such a match is. The system depends on a reference population of hundreds of human voices from which to learn what is the norm.
For Nolan, while these ASRs are "an extra tool" in the specialists' tool box, " there is a real danger that these systems hide from a jury the implications of the "complexities of the human voice and language".
However, while French acknowledges that they are "unlikely ever to do away with the human altogether", he argues that automated systems "are improving in their accuracy and objectivity and that there is some resistance to their use in the UK" that is preventing their wider adoption beyond labs such as French Associates.
"Systems such as Batvox provide centre-stage forensic evidence in court, even if just like other forensic tools it's not 'beyond reasonable doubt'. It can get near-100 per cent accurate at the moment depending on the quality and amount of speech input." Even fingerprinting, he adds, "isn't as reliable as it is portrayed on CSI." Similarly, Moreno feels the technology is improving and the reliability of other forensics tools is overstated. "Fingerprints are great in the lab but in the actual crime scene are often blurred or incomplete. Phone calls now are pretty stable," he says. "So it's comparable in its accuracy to many other tools, although the main problem is that at the moment there aren't enough voices in the databases." Some financial institutions are now building databases of suspect fraudsters' voices.
Perhaps the main problem these systems face is, apart from background noise, the length of the call. "Blackmail or kidnapping calls may only last five or six secs," Moreno says. "And you need six or seven seconds for an accurate result. Although even in these cases if the system can identify to three or four other kidnappings then it is a great help to the police."
In the end, for French, even though many engineers believe they can reach the holy grail of fingerprinting the human voice, "if it turns out that people's vocal tracts and speech-producing organs don't differ enough biologically from each other then there will be a limit to how accurate systems will be".
Nolan feels more strongly that "an individual doesn't have a voice, but many voices" so that a human specialist is always going to be needed to make a judgement.Reuse content