A name only when the evidence agrees
The speaker is the point of a quote: “our CEO said” and “someone mentioned our CEO” are different news. So we would rather say we’re unsure than guess.
Every quote says which it is
Colour never carries it alone: the state is always written out beside the quote.
We name the speaker, because the evidence for who said it agrees, as set out below.
We heard this in the episode but couldn’t tell for sure who said it. Trackers hold these back unless you ask for them too.
We found it’s better to teach the model some basic principles like what it should prioritize, how it should reason about what the right thing to do is.
Where one voice stops and the next starts
Before we can name anyone, we split an episode into turns: stretches of speech by one voice.
Shows with their own transcript
Some shows publish a transcript that says who is speaking. We use their labels as they give them.
One host and one guest
We read the words with their times and mark where the speaker changes. A sentence next to a change we aren’t sure of gets its own turn, flagged unsure.
Three voices or more
Panels and round tables go through a speech model that tells voices apart. Its labels, Speaker 1 and Speaker 2, mean nothing outside that episode.
Who each voice is
Then we look for evidence of who each voice belongs to, in the episode and in what’s written about it.
The kinds of evidence
- The show’s own transcript names who is speaking.
- The host: the show’s known host, when that voice is the one asking the questions.
- The guest list: someone named in the episode’s title or description.
- Themselves: someone says “I’m…” or “my name is…”.
- Someone else: another voice says their name to them.
What “confirmed” needs
The show’s own transcript names them, or two different kinds of evidence agree and none disagree. A host asking the questions who is also named in the description is confirmed. A name that only one kind of evidence suggests is not.
Everything else is unconfirmed
So is every sentence from a turn we flagged unsure in step one. We keep the evidence behind each name, so a decision can be checked.
We never name anyone from their voice
From what’s said and written
Names come only from the evidence above: transcripts, titles, descriptions, introductions and names said aloud.
Nothing about a voice is kept
The labels that tell voices apart exist only while an episode is processed, and we never match a voice from one episode to another.
No accuracy figure yet
We don’t publish a figure for how often we name speakers rightly, because we haven’t finished measuring one we’d stand behind.
Transcribing every new episode, with each speaker named this way, is in pilot. If we got a name wrong, write to hello@poddata.ai or use the takedown form, and we’ll fix it.
Hear when they speak, not when they’re mentioned.
Trackers send a quote only when the person you follow is the one speaking. They’re in pilot, by request.
Request pilot access