Speaker diarisation separates audio into who-said-what segments, and timestamping anchors each line to a precise moment — together answering who said it and when. Most engines re-tag speakers each call or lose identity after a pause. Oak keeps speaker identity consistent across meetings and timestamps every line to the second, so anyone can jump to the audio behind a quote.
Most teams ask “is the transcript accurate?” before they ask “who said it, and when?” That second question is the one that determines whether a transcript can stand up to a client dispute, an audit, or a journalist’s fact-check. Accuracy gets you a readable record; diarisation and timestamps get you a defensible one.
Key takeaways
- Diarisation answers "who," timestamps answer "when." They are separate jobs from transcription, and both are needed for a transcript to hold up under scrutiny.
- Diarisation is a different model from transcription. It ignores words and tracks voiceprint changes, handling overlaps, interjections, and similar-sounding voices.
- Most engines lose identity after a pause. A speaker who goes quiet and re-enters often gets tagged as a new person, fragmenting the record.
- Persistent, consent-based speaker memory fixes it. Oak learns your team's voices, keeps them tagged across meetings, and tags new participants as a Speaker until you confirm who they are.
- Timestamps to the second make a transcript evidence. When someone says "I never said that," you jump to the exact second in the audio — no scrolling, no manual cueing.
What is speaker diarisation?
Speaker diarisation is the task of separating audio into who-said-what segments. It runs alongside transcription but is a separate model: it doesn’t care about words, only about voiceprint changes — the acoustic signature that distinguishes one speaker from another. A good diariser handles overlapping speakers, brief interjections, and identical-sounding voices recorded on different devices. The output is what lets a transcript attribute each line to a named person rather than to an anonymous “Speaker 1.”
Why most engines get speaker identity wrong
Two failures are common, and both undermine trust in the record:
- Cross-call identity leakage. Some diarisers assign the same identity across calls without permission, which effectively shares voiceprint data between sessions that should be isolated.
- Identity loss after silence. When a speaker is quiet for more than a minute and re-enters the call, many engines tag them as a brand-new person — so one participant ends up split across two or three labels, and the record fragments.
How Oak handles speaker identification
Oak’s persistent speaker model (currently called Voice Memory) is workspace-scoped: it learns the voices of your team, with consent, and keeps them tagged across meetings. There is no re-tagging on every call, so a recurring participant is recognised session to session. When a new participant joins, they are tagged as a Speaker until you confirm who they are. Crucially, voiceprints stay inside your workspace and are never used to train shared models — the speaker identity that makes your transcripts useful never becomes someone else’s training data. This is the same isolation principle behind the workspace-scoped Jargon Library.
Why citation-grade timestamps matter
Every line in an Oak transcript is timestamped to the second. When someone disputes “I never said that,” the partner, journalist, or compliance officer jumps to the audio at that exact second — no scrolling, no manual cueing. This is the mechanism that moves a transcript from “rough notes” to “evidence,” and it is why timestamps are not a cosmetic feature but the backbone of an audit trail. For bilingual proceedings, where the official record needs both languages intact, the same second-level anchoring lets a reviewer verify any line against the source audio.
Diarisation and timestamps: generic vs Oak
| Capability | Oak | Typical generic engine |
|---|---|---|
| Speaker identity across meetings | Learned once, kept consistent across meetings (with consent) | Re-tagged each call as Speaker 1, 2, 3… |
| Speaker returns after a pause | Re-matched to the same identity | Often tagged as a new person |
| New participant | Tagged as a Speaker until you confirm | Silently assigned an identity |
| Voiceprint data | Stays in your workspace; never trains shared models | May be shared across sessions or used to train shared models |
| Timestamp precision | To the second, on every line | Approximate, paragraph-level |
| Dispute resolution | Jump to the exact second behind any line | Manual scrolling to find the moment |
Where this shows up
Most visibly in Legal and Statutory deployments, where the audio trail has to hold up — minutes become part of a case or public record, and a disputed line has to be verifiable against the source in seconds.
Oak for Legal Sector
The customer-facing deployment that uses the workflow described in this article.
Frequently asked questions
What is the difference between transcription and diarisation?
Transcription turns speech into text — the words. Diarisation separates the audio into who-said-what segments — the speakers. They are separate models: transcription cares about words, diarisation cares only about voiceprint changes. A transcript needs both to attribute each line to a named person rather than an anonymous "Speaker 1."
Why does my transcription tool keep relabelling the same speaker?
Because many diarisers lose a speaker's identity after a pause. When someone is silent for more than a minute and rejoins, the engine tags them as a new person, so one participant ends up split across several labels. Oak avoids this by keeping a consistent, consent-based speaker identity across the meeting and across sessions.
How does Oak keep speaker identities consistent across meetings?
With a workspace-scoped speaker model that learns your team's voices, with consent, and keeps them tagged from one meeting to the next — no re-tagging each call. New participants are tagged as a Speaker until you confirm them, and voiceprints stay inside your workspace and are never used to train shared models.
What does "citation-grade" or "to the second" timestamping mean?
It means every line of the transcript is anchored to a precise second in the recording, so you can click a line and jump straight to that moment in the audio. When a statement is disputed, a reviewer verifies it against the source in seconds, with no scrolling or manual cueing — which is what makes the transcript usable as evidence.
Are timestamps and speaker labels enough for a legal or audit record?
They are the foundation of one. Second-level timestamps plus consistent speaker attribution let a partner, auditor, or compliance officer trace any line back to the source audio — which is what lets a transcript serve as evidence rather than just notes. A fully audit-ready process also depends on your own retention and access policies, which should be confirmed against your regulatory obligations.