Attribution Trust
How we decide which words on this site are spoken by which person, what we publish, what we hold back, and how you can correct us.
Speaker-attribution policy v1.0
Attributed quotes published
44088084
Suppressed by human review
0
Human-verified voice samples
0
Reader reports open
0
How attribution works
Every podcast episode runs through automatic speech recognition, speaker diarization (separating who-speaks-when), and a voice-identity matcher. The matcher compares each speaker's voice against a library of voice profiles built from prior identified appearances, and assigns a person only when the voice similarity clears a publication threshold.
On top of the matcher sits a layered review system: automated safeguards that withhold attributions wholesale when risk patterns appear (for example, implausibly dense attribution of one voice across a show, or a language mismatch between a voice profile and an episode), human reviewers working from a moderation queue, and reader reports. Together these have withheld millions of machine-generated attributions from public display — holding back when uncertain is a feature of this platform, not an afterthought.
What the badges mean
Quotes attributed under policy v1.0 carry a badge keyed to the raw voice-similarity score of the match (cosine similarity between voice embeddings, on a 0–1 scale).
| Badge | Meaning |
|---|---|
| voice-verified | Voice similarity of at least 0.95 against the person's profile — our highest automatic tier. |
| podcast-host | Matched as a declared host of this show. Host voices have the most samples and the most scrutiny. |
| human-verified | A human reviewer has confirmed this attribution. |
| matched | Voice similarity above our publication threshold but below the voice-verified tier. |
| low-confidence | Below current thresholds. Mostly hidden from public view; shown only where context clearly demands it. |
Quotes attributed before policy v1.0 render without a badge — we did not record raw similarity scores for them, and we won't invent one retroactively. They remain subject to the same suppression system, are progressively re-reviewed, and can be reported like any other quote.
What gets published
Every attribution carries a status — published,
unreviewed, or
suppressed. Only published rows appear
on public pages. The gate is enforced at render time, not just at search
time, so even a deep link to a suppressed quote will refuse to display it.
We suppress when the matcher is uncertain, when attribution density on a single show suggests the diarization step bucketed multiple voices together, when the language of an episode and a voice profile disagree, when a human reviewer catches a mistake, or when a credible reader report comes in. Suppression decisions are recorded in an internal audit log with reviewer, reason, and timestamp.
If we got it wrong
Every attributed quote on a person page has a "Not <name>?" link next to it. One click queues the quote for human review. We do not auto-remove on report — a single click should not be able to take down content — but every report lands in front of a human moderator.
For sensitive or legal corrections, write to [email protected]. We aim to respond to good-faith correction requests within five business days.
What we don't claim
- We don't claim 100% accuracy. We claim layered safeguards, a measurable process, and a transparent path to correction.
- We don't claim a quote captures a person's full meaning out of context — these are transcribed audio segments, not endorsed positions.
- We don't infer beliefs, intentions, or relationships from quotes. Other parts of the platform make connections between entities, and those have their own review policy.
- Attribution may change. A quote published today may be suppressed tomorrow if review finds the matcher made a mistake — and occasionally the reverse.