Home

The Research Hub for Language in Forensic Evidence was established to address injustice arising from the legal handling of indistinct speech recordings admitted as forensic evidence in criminal trials.

The main sources of injustice the Hub has identified in the use of forensic speech evidence are

  • use of unreliable transcripts and translations to assist the court in understanding indistinct audio recordings
  • admission of audio that has been ‘enhanced’ with unreliable processes, and
  • unreliable attribution of utterances in audio recordings to specific speakers, in English and other languages.
Read more about this video on our FAQ page

The Hub’s overarching goal

To build the research base needed to ensure that all audio used as evidence in criminal trials is

  • processed in accountable ways that have been demonstrated not to alter how the content is understood
  • accompanied by a reliable transcript, produced via accountable, evidence-based methods with reliable translation and speaker attribution where needed, and
  • provided to jurors in ways that enable them to reach a reliable interpretation of the content of the audio evidence.

Our innovative approach

It is common to assume that forensic transcription research would focus on analysing the audio itself – and of course audio analysis is important. However long-standing scientific findings indicate that, with indistinct recordings, the listener is at least as important as the audio. This is shown by the fact that the same audio can be heard in dramatically different ways by different listeners, or under different listening conditions.

In its travels through the legal process, forensic audio is heard by multiple listeners, including the transcriber, and, crucially, the lawyers, judges and jurors evaluating the transcript later, in court. We have to be sure that all these listeners understand the content correctly.

For that reason, the Hub’s approach draws on scientific knowledge about speech perception, an interdisciplinary field incorporating findings from speech science, linguistic science, and various branches of psychological and social science.

Finally, and perhaps most importantly, the Hub bases its methods on detailed understanding of exactly how forensic audio and transcripts are handled in the legal process, not just in open court, but also behind the scenes.

Current projects

Establishing accountable methods for producing demonstrably reliable transcripts of forensic audio, whether in English (including non-mainstream varieties) or in other languages

This project runs experiments using human transcribers, rigorously tested for aptitude in deciphering indistinct audio, and a bespoke transcription platform called ‘SoundScribe’, custom-built for the Hub by Melbourne Data Analytics Platform. SoundScribe has been designed to collect transcripts from listeners, and enable expert analysts to compare and evaluate multiple transcripts produced under different conditions.

Think you might have aptitude for deciphering indistinct audio? Interested to be on our experiment participant list? Contact us – we’d like to hear from you!

Disseminating reliable information about the capabilities and limitations of forensic speech enhancement and automatic speech recognition, especially in relation to methods that use AI or machine learning

This project is developing a database of forensic-like audio, and testing a range of cutting-edge enhancing techniques to determine the extent to which they help listeners decipher forensic-like audio with and without reliable or unreliable transcripts. The results are already enabling the Hub to overcome many common misconceptions about forensic speech enhancing that affect the legal process.

Providing guidance on the best ways to present audio and transcripts to juries so as to ensure they reach a reliable understanding of the audio content

With forensic speech recordings, it is not enough for an expert process to reach a reliable conclusion about a transcript or an enhancement. The Hub aims for an end-to-end solution that ensures the court as a whole, and especially the jury, reach a reliable understanding of the audio content.

Finding effective ways to counter misconceptions about speech, speech perception and transcription that affect the legal process

One of the Hub’s main findings is that injustices with forensic audio arise from legal procedures developed on the basis of misconceptions about the nature of speech, how human speech perception works, and how speech can be represented with a transcript.

Through detailed case studies and close collaboration with academic and practising lawyers, the Hub has developed effective demonstrations to communicate these misconceptions and their effects on the legal process to the judiciary.

Theoretical implications of forensic transcription and enhancing

The Hub’s projects need practical experimentation, but they also require advances in theoretical understanding of how speech is perceived, and how what is perceived is represented in a transcript. The Hub approach departs from standard computational models of perception and cognition (useful as these can be for their own purposes).

Instead we understand speech perception to be a Bayesian process, in which listeners start with vague prior expectations about the kinds of words they are likely to hear, then seek confirmation in the incoming data – updating their expectation as needed until they arrive at a conclusion which they accept as their perception.

Contact us

For more information on any aspect of the Hub’s work, including casework requests, please contact the Director, Prof Helen Fraser.

For general enquiries, please email the School of Languages and Linguistics: soll-info@unimelb.edu.au