# Voice Identification for Lawyers: What a Similarity Score Can (and Can't) Prove
"Who is speaking?" is often the most important question in an audio recording. Witness statements, threats, admissions — all depend on knowing who said what.
Automatic speaker diarization answers "how many speakers?" and "when did they switch?" It does not answer "which speaker is which person?"
Voice identification tools answer "does this segment sound like the enrolled voice?" — but they do not testify. A similarity score is corroboration, not proof.
What speaker diarization does
Diarization separates a multi-speaker recording into segments and labels them: SPEAKER_00, SPEAKER_01, SPEAKER_02.
It is based on acoustic similarity. Segments that sound alike are grouped. Segments that sound different are separated.
Diarization does not know who the speakers are. It only knows they are different.
This is a common issue in jail call transcription, where automatic systems separate speakers but cannot name them without additional context.
What voice identification adds
Voice identification (also called speaker recognition or voice biometrics) compares an unknown segment to a known reference sample.
You provide a reference recording of a known speaker — a prior call, a deposition, a known voicemail. The system generates an embedding (a numerical representation of vocal characteristics) and scores every segment in the disputed recording against it.
The result is a similarity score: 0.0 (no match) to 1.0 (very high similarity).
What a similarity score means
A high similarity score (e.g., 0.85) means the acoustic characteristics of the unknown segment are consistent with the enrolled voice.
That is corroboration. It supports your hypothesis that Speaker A is the enrolled person.
It is not identification. The system does not know who the person is. It only knows the segments sound similar.
What a similarity score does not mean
It is not a unique identifier. Voices are not fingerprints. People with similar vocal characteristics can produce similar embeddings.
It does not account for recording conditions. A phone call and a bodycam recording of the same person may score lower than expected due to codec differences, noise, or microphone quality.
It does not prove identity in court. A machine cannot testify. The score is a tool for a human expert or fact-finder to use — not a verdict.
When voice identification works well
Voice identification is most reliable when:
- The reference sample and the disputed recording are from similar conditions (same phone system, similar quality)
- The reference sample is clean and representative (at least 30 seconds of continuous speech)
- The disputed segments are clear enough for feature extraction
- The enrollment is from the same time period (voices change with age, health, stress)
When it fails
Voice identification fails when:
- The reference sample is too short or too noisy
- The disputed audio is severely degraded (heavy compression, wind, distortion)
- The speaker is deliberately disguising their voice
- There is significant background noise or overlap
When the system cannot score reliably, it should say so — not guess.
The "can't confirm" result is valid
If a segment scores 0.45 — neither high nor low — the honest answer is "inconclusive."
That is not a failure. It is a constraint. The audio quality or the speaker's variability makes confident scoring impossible.
Some tools will force a binary yes/no. That is a design choice, not a fact. A tool that admits "I can't confirm" is more trustworthy than one that always guesses.
How VeriVox handles voice identification
VeriVox allows you to enroll a reference sample of a known speaker. Every segment in the recording is scored against that reference.
High similarity (>0.75) — segment is likely the enrolled speaker
Moderate similarity (0.50–0.75) — inconclusive, flag for review
Low similarity (<0.50) — segment is likely not the enrolled speaker
Scores are logged per segment. If the diarization labeled a segment as SPEAKER_00 but it scores high against the reference for SPEAKER_01, that conflict is flagged.
The human reviews the flagged sections and certifies the attribution. The export labels which segments were machine-attributed and which were human-certified.
Enrollment method matters
The quality of the reference sample affects every downstream score. Best practices for enrollment:
- Use clean audio — minimal background noise, clear speech
- At least 30 seconds — longer is better; more speech = better embedding
- Representative speech — conversational tone, not shouting or whispering
- Same language and dialect — embeddings are language-specific
If the reference sample is a 5-second voicemail recorded in a moving car, the scores will be unreliable.
Multiple enrollments
If you have reference samples from multiple known speakers, enroll all of them. The system scores every segment against every enrolled voice.
The output is a matrix: each segment scored against each reference. That reveals:
- Segments that match one voice strongly and others weakly (clear attribution)
- Segments that match multiple voices moderately (inconclusive)
- Segments that match no enrolled voice (unknown speaker or poor audio)
Conflicts between diarization and identification
Automatic diarization sometimes swaps speakers — especially when:
- Two speakers have similar voices
- Audio quality degrades mid-recording
- There is a long silence before a speaker returns
Voice identification catches those swaps. If diarization says SPEAKER_00 but the segment scores high for the enrolled SPEAKER_01, that is a conflict.
VeriVox flags those conflicts and queues them for human review.
Legal use of voice similarity scores
Voice similarity scores are used in:
- Investigative work — narrowing suspects, corroborating hypotheses
- Litigation support — organizing multi-speaker recordings, flagging contested lines
- Expert reports — forensic voice examiners use embeddings as part of a broader analysis
They are not typically admitted as standalone proof of identity. They support testimony.
What an expert witness would say
A qualified forensic voice examiner testifies to:
- The method used to compare voices
- The quality of the reference sample and disputed audio
- The similarity score or categorical conclusion (match, likely match, inconclusive, exclusion)
- The limitations of the method
- Their opinion, based on the analysis and their experience
The similarity score is one input. The expert's judgment is the conclusion.
What you can do without an expert
If you are not presenting expert testimony on voice identification, you can still use similarity scoring as an organizational tool:
- Enroll known speakers
- Score segments to corroborate diarization
- Flag conflicts for your own review
- Organize excerpts by attributed speaker
The scores do not go into evidence. They help you prepare the case.
VeriVox's speaker verification pipeline
VeriVox's Verify stage:
- You enroll reference samples of known speakers
- Every segment is scored against every enrolled voice
- Conflicts with diarization are flagged
- Segments are queued for human review, sorted by importance
- You listen and certify the attribution
- The export labels each segment: machine-attributed or human-certified
The similarity scores are preserved in the export metadata. If an expert later reviews the file, they have access to the raw scores.
For investigators using voice identification to organize multi-speaker recordings, learn more about VeriVox for investigative work →
Checklist: Using voice identification responsibly
- [ ] Use a clean, representative reference sample (≥30 seconds)
- [ ] Enroll all known speakers if possible
- [ ] Score segments and flag conflicts with diarization
- [ ] Review flagged segments by ear
- [ ] Treat scores as corroboration, not proof
- [ ] Label attribution as machine or human-certified
- [ ] If presenting scores in court, consult a forensic voice expert
FAQ
Can I enroll a voice from a different recording than the one I'm analyzing?
Yes — that is the typical use case. The reference sample can be from a known voicemail, deposition, prior call, or any clean recording of the person.
What if I don't have a reference sample?
Then you cannot use voice identification. Diarization will still separate speakers, but you will need to attribute them manually by listening.
How accurate is voice identification?
Accuracy depends on audio quality, enrollment quality, and the similarity of speakers. In controlled conditions (clean audio, good reference), modern systems are highly reliable. In poor conditions, they are not.
Can someone fake a voice and fool the system?
Professional impersonation or voice synthesis can produce high similarity scores. Voice identification is not foolproof. It is one form of evidence, not the only form.
Can I use this to identify an unknown speaker?
No. Voice identification compares an unknown sample to a known reference. If you have no reference, the system cannot identify the person. It can only say "these segments sound like the same speaker."
What happens when voice identification disagrees with the transcript?
If a segment is attributed to Speaker A by diarization but scores high for enrolled Speaker B, that conflict is flagged for review. Handling disagreement in transcripts — whether between models or between attribution methods — is part of producing defensible evidence.