Forensic voice comparison

The questioned recording contains the voice whose identity is disputed. The reference, or known-speaker, recording contains speech attributed to an established speaker. Morrison and Enzinger's introduction to forensic voice comparison describes the task as helping assess whether the recorded voices came from the same speaker or different speakers.

Keep the question separate from neighboring work. Transcription concerns the words. Restoration concerns making an existing recording more useful to hear. An authenticity examination concerns the recording's claimed origin or history. A speaker comparison evaluates voice evidence relating to identity. A file can raise more than one of these questions, and answering one does not settle the others.

An unknown recording on its own is not the two-source comparison described here. Establish what reference material exists before expecting a comparison report. The guide to restoration, transcription and authentication can help you describe the recording question clearly.

Forensic voice comparison needs suitable questioned recordings, which contain speech whose speaker is disputed, and reference recordings whose speaker's identity is established independently. Provide the best available originals and explain where each recording came from. Identify the speaker and passages to be examined, while keeping the reference speaker's identity separate from the conclusion you hope the analysis will support.

Tell the examiner what reference recordings are available, rather than selecting only the clip that sounds most similar to the questioned voice. Agree which originals are suitable to share for the assessment, including any permission or handling requirements.

Ask the examiner to assess the amount of usable speech, its variety, and the features that can be compared. File length alone is a poor description: a long recording may contain little speech from the relevant person, while several passages may repeat a narrow set of sounds. Morrison and Enzinger discuss both short recordings and adverse recording conditions as practical obstacles.

Avoid treating a universal minimum duration as a guarantee of suitability. The examiner needs to consider the actual speech and proposed method. If there are several questioned clips, do not assume they all contain the same speaker. Their relationship is itself something that may need checking.

A speaker's voice changes between occasions. Conversation, reading aloud, illness, emotion and effort can affect its characteristics. Different people can also share features such as a regional pronunciation or similar pitch. The statistical-models chapter by Morrison and colleagues distinguishes variation within a speaker from variation between speakers.

For an evaluation using a likelihood-ratio framework, similarity is only part of the question. Typicality matters too: how common are those features among the relevant alternative speakers? Morrison and Enzinger explain why a resemblance can carry different weight depending on that comparison population.

Here, the relevant population represents plausible alternative speakers under the different-speaker proposition. Ask why the data used to represent that population are appropriate to the case.

A shared feature therefore needs context. Matching one pitch measurement, spotting a similar spectrogram pattern, or getting a high similarity score does not by itself establish identity. Ask which features were evaluated, which alternative speakers the evaluation represents, and how the method accounts for variation.

A phone call and a room recording capture more than the speaker's voice. Microphones, distance, reverberation, transmission and compression can affect the recorded signal. Speaking style may differ as well. Morrison and colleagues describe these as sources of mismatch between questioned and known-speaker material.

List the differences you know about. Was one passage whispered and the other spoken normally? Did one file pass through a messaging service? Were the recordings made years apart? These details help the examiner decide which comparisons and validation conditions are relevant.

Keep any processed version alongside its original and record what changed. A cleaner-sounding copy does not demonstrate that the features used for speaker comparison have been preserved. Discuss processing with the comparison examiner rather than assuming that stronger noise removal produces stronger identity evidence.

Different languages can change the speech sounds and patterns available for comparison. Code-switching adds another consideration when a person changes languages within the material. The examiner needs to identify comparable features and explain the limits caused by language differences.

ENFSI's speaker-comparison manual calls for particular caution with language mismatch and appropriate language expertise. A translated transcript can help explain the meaning of a conversation, but it does not turn the original speech into a same-language acoustic sample.

Provide the original audio, identify the languages when known, and label translated text separately. Ask how the selected method was assessed under the language conditions involved. Avoid assuming that a multilingual system has been validated for every language pair or recording situation.

Ask for evidence that the method was tested under conditions relevant to the material. This validation checks performance using suitable test recordings. Useful questions include whether the validation includes comparable recording channels, speaking conditions and relevant speakers, and whether the test material is independent of the material used to develop the system. Morrison and colleagues discuss case-relevant validation and testing of automatic systems.

Forensic voice comparison methods may use auditory listening, acoustic-phonetic analysis of speech features or human-supervised automatic systems, sometimes in combination. These describe how information is extracted; they do not alone tell you how reliable a particular conclusion will be. Morrison and Enzinger distinguish the extraction approach from the framework used to evaluate its findings.

Acoustic features can include fundamental frequency, which relates to pitch, and vowel resonances, the frequency patterns shaped by the vocal tract. Automatic speaker recognition can compare speaker information extracted from recordings. Its technical output still needs interpretation within the forensic evaluation framework; an individual feature or raw score is not an identity conclusion.

Calibration concerns how system scores are related to the reported strength of evidence. A raw similarity score and a calibrated likelihood ratio—a measure of support for stated same-speaker and different-speaker explanations—are different outputs. A published benchmark or advertised accuracy percentage needs its test conditions before it can help you assess a case.

Ask whether the testing covers the complete procedure used to reach the reported conclusion, including calibration and any combination with human assessments. Testing one automatic component does not, by itself, establish how the combined procedure performs.

Human decisions also remain part of the work, including selecting material and interpreting limitations. Ask how irrelevant contextual information is controlled, what the report documents, and how the analysis is reviewed. ENFSI's manual discusses documentation, cognitive bias and review by another expert.

Read the propositions before interpreting the result. In a comparison that places the same-speaker proposition in the numerator, a likelihood ratio compares:

The likelihood of the observed voice evidence under the same-speaker proposition, divided by its likelihood under the stated different-speaker proposition and relevant population.

Under that orientation, a value above one supports the same-speaker proposition relative to the alternative; a value below one supports the different-speaker proposition; one does not favor either proposition. The assumptions and relevant population remain part of the interpretation. ENFSI's evaluative-reporting guideline explains this framework as evaluating support for competing propositions.

This is not the probability that the person spoke, nor the probability of guilt. Those are different questions. A numerical result should not be rewritten as a percentage identity match. Reports can also use verbal conclusions. The Forensic Voice Centre’s explainer describes its verbal scale, while ENFSI notes that scales vary between laboratories and jurisdictions. Check the scale and explanation in the actual report.

Similarity does not resolve every authenticity concern either. Univaso and San Segundo's 2025 study found limitations in the tested cloned-voice detection methods across different algorithms and degraded recordings. Those bounded results are a reason to ask separate questions about synthetic speech when relevant, rather than treating one similarity or detector score as a complete answer.

Before arranging work, assemble the original recordings, the passages of interest, available dates, recording-device or application details, and the copying or processing history. Explain any gaps in that history. The evidence-preservation guide covers keeping originals and working copies distinct.

Confirm the proposed comparison, the examiner's relevant qualifications, the suitability assessment, the validation evidence and the intended report. Ask what would make the result limited or unsuitable. Confirm what the comparison quote includes, which report or review deliverables it covers, and when the examiner can provide them. Discuss permission and file handling before sharing identifying material.

If the immediate problem is hearing the recording, a restoration review addresses that listening job. Media Medic's engineers listen before invoicing for restoration work. If your question is who spoke, confirm the specific comparison scope before arranging work; restoration and recording authentication do not themselves establish speaker identity.

This guide compares published explanations of forensic speaker comparison, professional methodology and evaluation guidance, and research on recording and synthetic-speech limitations. It distinguishes the material being compared, the method used, and the meaning of the reported conclusion.

It does not conduct a voice comparison, calculate an accuracy rate, evaluate a case or state that evidence will be admitted in a particular court. Methods and reporting practices vary. The relevant examiner must explain the approach, validation and limits for the actual recordings.

A useful conclusion makes its question, material, assumptions and limits clear. Start with suitable original and reference recordings, ask how mismatch and validation were handled, and read the result as the stated strength of evidence rather than an identity guarantee.