“Our CEO said” and “someone mentioned our CEO” are different news. A tool that mixes them up sends the wrong alert. So the speaker is the part of a quote we work hardest on.
Two states, always written out
Every quote we show says which state it’s in.
- Speaker confirmed. We name the speaker, because the evidence for who said it agrees.
- Speaker unconfirmed. We heard it in the episode, but couldn’t tell for sure who said it.
We’d rather say we’re unsure than guess. Trackers hold back unconfirmed quotes 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.
Step one is splitting the audio by voice
Before we can name anyone, we split an episode into turns. A turn is a stretch of speech by one voice. People who build speech tools call this speaker diarization.
- Shows with their own transcript. Some publish one that says who is speaking. We use their labels as given.
- 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 doubt gets its own turn, flagged unsure.
- Three voices or more. Panels go through a speech model that tells voices apart. Its labels mean nothing outside that episode.
Step two is finding who each voice is
Then we look for evidence of who each voice belongs to. There are five kinds.
- The show’s own transcript names who is speaking.
- The host is the voice asking the questions.
- The guest is named in the episode’s title or description.
- Someone says their own name.
- Another voice says their name to them.
A speaker is confirmed when the show’s own transcript names them. They’re also confirmed when two different kinds of evidence agree and none disagree. A name that only one kind suggests stays unconfirmed. We keep the evidence behind each name, so any decision can be checked.
Why we never use voice prints
Speaker identification is often done by matching a voice to a stored sample. We don’t do that. Names come only from what is said and written about an episode.
The labels that tell voices apart exist only while an episode is processed. We never match a voice from one episode to another. Nothing about a voice is kept.
What we don’t claim
We don’t publish a figure for how often we name speakers rightly. We haven’t finished measuring one we’d stand behind. If we got a name wrong, write to hello@poddata.ai and we’ll fix it.
The full rules are on how we name speakers.