In this article

What does the annotation leave unexplained?

Imagine an examiner pausing over a short ridge near the edge of an impression. They record its position, then explain that pressure may have altered its appearance and that another area will be needed to check the interpretation. Later, they revise the feature after examining the reference.

The saved annotation identifies an observation. The explanation adds the tentative account, the proposed check and the reason for returning to it. A final feature label may contain little of that development. Commentary offers a way to retain some of the explanation alongside the visual record.

That possibility also creates a research question: what does an examiner actually verbalise, and how should another person interpret it? A fluent account can be useful evidence about an examination without being a complete description of the processes that produced the decision.

Research can examine the content of an explanation

Corbett and Tangen used concurrent think-aloud protocols with 44 qualified fingerprint examiners and 44 matched novices. Each participant completed six comparisons while verbalising their thoughts. The researchers transcribed and segmented the speech, then examined both the proportion of coded statements and their qualitative content. The account discussed here comes from the retained 2023 author preprint of their 2024 paper.[1]

The expert accounts included causal explanations for differences, checks on emerging interpretations and appraisal of the evidence. A reference to pressure, movement or distortion can explain why an examiner considers two appearances compatible. A check on a ridge relationship can reveal how the examiner tests an interpretation that already seems plausible.[1]

Validation and considering alternatives also have different purposes. Seeking another observation that supports an emerging account differs from looking for something that could overturn it. Both experts and novices discussed alternatives in this study; the authors found qualitative differences in the basis and depth of those explanations.[1]

Counting statements does not capture their meaning

The study makes a useful methodological distinction. The groups did not differ significantly in the proportions of planning or alternative-explanation codes after the study’s correction for multiple comparisons. Their statements could nevertheless differ in specificity, causal understanding and adaptability. Counting a category and interpreting what was said answer different questions.[1]

The authors also discuss recursive examination: returning to an earlier interpretation after seeing the known impression. They recommend documenting the adjustment and its substantive reason, such as a pressure or surface effect that led an examiner to reconsider a detail. The proposed value lies in making the change available for scrutiny; the study did not establish that requiring commentary improves examination accuracy.[1]

Expert explanations themselves need appraisal. Participants described different ways of judging significance and sufficiency. An examiner’s statement that a feature is rare or important records their assessment; it does not supply a verified statistical weight. The authors call for further work on terminology and evidence-based evaluation.[1]

Three sources with different roles

PROFAS distinguishes the original recording, the machine transcript and the wording accepted into the examination. Where the protocol enables dictation, an examiner records a clip, reviews the generated text and appends the reviewed wording to a field. The ordinary Record or Submit action then accepts that contribution with its speech-source link. A transcript sitting in a draft is still pending work.

A commentary-enabled protocol also allows an audio-backed account alongside the examination. The record can retain reviewed or typed wording with the original clip. This gives commentary a place beside observations and Evaluation while preserving the written information those stages require.

This separation matters because machine text can introduce meaning. Koenecke and colleagues’ Careless Whisper study identified phrases absent from the source audio in transcriptions of AphasiaBank recordings. The study examined Whisper in 2023, with speech and recording conditions different from forensic dictation. Its reported rates cannot be transferred to the transcription model used by PROFAS.[2]

The relevant design response is to retain the original and make review explicit. A disputed phrase can be checked against what was recorded. Researchers evaluating this workflow should distinguish ordinary transcription mistakes, unsupported additions and the examiner’s corrections, including whether a correction changes uncertainty, negation or the interpretation of a finding.[2]

The Inspector presents generated wording for review before it is appended to the field. This illustrative draft in the actual interface still awaits ordinary recording; the machine transcript remains separately available.

Timing and access need their own interpretation

A clip retains its examination context, with source-view information where attached, its stage and separate capture and acceptance timings. These help place an account within the record. They do not establish when the examiner first perceived a feature. The start of speech can follow observation, and acceptance can follow review of the wording.

Playback follows access to the accepted contribution. A verifier cannot use its linked audio to bypass withheld target findings. Researchers inspecting an earlier accepted record receive the speech links belonging to that basis, rather than later commentary silently added to an earlier account.

The protocol also determines whether the examination permits typed notes only, dictation or commentary. That distinction matters for a study: a short dictated note, continuous think-aloud instructions and a retrospective explanation are different recording conditions. Their contents should be interpreted with the instructions that produced them.

What should be investigated next?

Protocol analysis and transcription research encountered during the current project open further questions about the examination record. They suggest useful studies of what commentary adds, how another examiner understands it and how reviewing machine text affects the accepted account.

Such studies need consistent instructions, suitable recording conditions and attention to whether speaking changes the task itself. The protocol-analysis experiment used six comparisons per participant with limits on image manipulation, and verbal reports leave some internal processes unexpressed. Its findings do not establish a complete account of routine casework.[1]

Teaching examples and image-linked language datasets are further possibilities. Their use needs its own permissions, participant arrangements and research design. An expert’s explanation supplies material to investigate; its availability does not validate forensic model training. The immediate aim is more concrete: preserve an examiner’s recorded account with enough context for another person to inspect it.

FOLLOW THE EVIDENCE

Sources & further reading

The original research behind this article.

  1. Corbett, B. J., and Tangen, J. M. (2024). Capturing fingerprint expertise with protocol analysis. Applied Cognitive Psychology, 38(6), e70010.(opens in a new tab)

    Author preprint dated 20 December 2023; the journal article was published in 2024.

    Back to the article
  2. Koenecke, A., Choi, A. S. G., Mei, K. X., Schellmann, H., and Sloane, M. (2024). Careless Whisper: Speech-to-Text Hallucination Harms. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, 1672–1681.(opens in a new tab)

    Whisper experiments on AphasiaBank speech; not a PROFAS transcription benchmark.

    Back to the article