FAQ

Any kind of language can be admitted as an exhibit for the jury to consider in reaching their verdict. It just has to be captured in a way that lets it be observed and discussed in court. One way language evidence can be captured is in writing – traditionally in a paper document, but now often in electronic forms such as text messages or social media. However, the Hub is most interested in addressing problems with the use of spoken language as evidence in criminal trials.

Speech evidence is usually captured in an audio recording. Typically, the audio is obtained via covert (secret) surveillance during a police investigation, but increasingly it comes from sources like mobile phones or body-worn cameras. Sometimes the 'capturing device' is simply the memory of an 'ear witness', but it is now recognised that this kind of evidence can be very unreliable.

Speech recorded secretly (and legally) can allow the court to hear speakers making admissions or even confessions that they would not make openly. This provides extremely powerful evidence, capable of affecting the jury's interpretation of other evidence, and ultimately their verdict of guilty or not guilty.

The problem is that forensic audio is often of very poor quality, to the extent that listeners need the assistance of a transcript to hear what is said, and who is saying it. Of course, it is essential that any transcript used for this purpose should be reliable. Otherwise, it might 'assist' the court to hear words other than those that were originally spoken.

That’s why a major strand of the Hub's research is about 'forensic transcription', the science of how to provide demonstrably reliable transcripts of poor-quality audio, in English (including non-mainstream varieties) or in other languages.

Forensic transcription research does not take place in a vacuum. The law has its own long-standing procedures for handling poor-quality forensic audio – and these create a range of problems that are known to cause injustice.

The first problem is that the courts allow transcripts to be produced by, or edited by, detectives investigating the case. Of course, these police transcripts are not given directly to the jury. The law has safeguards intended to protect juries from being misled by inaccurate transcripts.

The deeper problem is that these legal safeguards are ineffective. They rely on lawyers, judges and juries checking the transcript against the audio. Long-standing research shows that the very act of checking can 'prime' listeners to hear in line with the transcript, even if it is thoroughly inaccurate.

That means that, rather than protecting juries from transcript errors, the legal safeguards can propagate errors to the whole court. Multiple case studies have demonstrated the serious injustices that can arise. It is important to note that these injustices can go both ways: innocent defendants can be found guilty, or guilty defendants set free.

Recent developments with large language models have brought amazing advances in audio enhancing and automatic transcription – for reasonably clear speech. Poor-quality forensic speech recordings take us into a different realm completely. While the Hub is carefully monitoring the latest audio enhancing techniques, so far we have found none that can consistently and reliably transform indistinct forensic-like audio into clearly intelligible speech. And so far, even the best current automatic speech recognition systems often fail with forensic-like audio.

For the time being, then, the Hub's methods rely on expert human listeners. We aim to develop tools that enable forensic audio to be processed and transcribed in responsible, evidence-based ways that assist with the end-to-end goal of ensuring that juries gain a reliable understanding of the spoken content of forensic recordings.

It is common to blame transcript injustice on police or others producing inaccurate transcripts; or on lawyers failing to detect significant inaccuracies. But careful analysis suggests a deeper cause.

Current legal procedures are founded on the concept that understanding spoken language is a matter of common knowledge, safely left to judges and juries. However, while understanding spoken language is certainly a common skill, common knowledge about how that skill operates has been shown to rest on a network of misconceptions.

One example is the misconception that responsible listeners, like judges and juries, can reliably evaluate a transcript of indistinct audio simply by checking it carefully against the recording. Another is the misconception that responsible listeners can easily determine if ‘enhancing’ has made audio ‘clearer’.

The Hub argues that the way to address this source of injustice is for the law to treat forensic transcription as a science. That way audio can be transcribed via accountable, evidence-based practices like those used by other responsible forensic sciences, not as a matter of common knowledge, with all the misconceptions that involves.

Making this change is a challenge for the law, and the Hub is in active discussion with judges and legal scholars, helping to discover how it can best be brought about. But forensic transcription also poses interesting challenges for linguistic science.

Linguistic science has a great deal of relevant knowledge, collected over many decades, to explain why current legal procedures are bound to create injustice. However, to this date, our field has not developed a set of evidence-based methods specifically designed for creating demonstrably reliable transcripts of indistinct forensic audio.

That is one of the projects the Hub is currently engaged in; developing accreditation tests to identify listeners with high-level aptitude for deciphering poor-quality audio, and rigorous, evidence-based methods that enable them to produce either a demonstrably reliable transcript, or a demonstrably reliable opinion that the audio is not capable of transcription to a level of certainty suitable for use in court.

Research findings about forensic audio challenge current theories that see human speech perception as a kind of computation, so the Hub looks in new theoretical directions. We are especially interested in developing Bayesian theories that see the human listener as a purposeful, meaning-seeking agent, actively updating hypotheses about the words that are contained in incoming speech.