You read a transcript and something is off, but when you go back through it, every word is right, or near enough. The problem is who the words belong to: two people have been run together into one speaker, one person has become three, or a whole exchange has been attributed to another person who wasn’t present.
The words are right and the transcript is still wrong. A record of an interview, a hearing or a recorded statement exists to show who said what. When a line is attributed to the wrong person, the record has them saying something they never said. A wrong speaker is also expensive to put right. When two people are merged into one speaker, the error runs for as long as they keep talking, and untangling it means going back to the recording to hear who was actually speaking. Sorting out the speakers is a large part of correcting an automated draft.
What speaker identification is
Getting the right names on a transcript is two jobs, separate from getting the words right. The first is splitting the recording by speaker: deciding where one voice stops and another starts, and which stretches belong to the same voice. The engineering term for this is diarization. The second is putting names to the voices, which the recording alone often cannot supply.
Why speakers get mixed up
- People talk over each other. A microphone records one signal. When two people speak at once, both voices are in it together, and the system has to pull them apart after the fact. Most systems assume one person speaks at a time, so overlap is where they most often drop words or give them to the wrong person. Interruptions, objections and corrections are exactly the moments that overlap.
- Voices sound alike. Two speakers with similar pitch, accent and pace often get merged into one. Two men on a panel, two sisters in an interview and colleagues who have worked together for years are common cases.
- The microphone is far away, or the line is poor. A speakerphone, a conference room, a phone line and a room with an echo all blur what tells one voice from another, and background noise can be mistaken for another speaker.
- Turns are short. A “yes,” an “objection” or a name said once is over before the system has heard enough of the voice to place it.
- Nobody said who was speaking. A system can tell that a second voice is present. It cannot know that the second voice is the witness unless someone says so on the recording or supplies the names afterward.
What you can do before a vendor is involved
The biggest improvements happen at the recording, not at the transcript.
- Record each participant on their own channel where you can. A courtroom or interview room system with a microphone per position keeps the voices separate at the source. A single phone recording of a group call is the opposite case, and no later processing fully recovers it.
- Put the names on the record. State who is present at the start, and have people identify themselves when they begin speaking. The recording then carries what the system cannot guess.
- Leave a short pause between speakers. Have the person answering pause before they reply, and the person asking pause before the next question. The gap marks where one voice ends and the next begins, so a one-word answer is not run into the question around it.
For people talking over one another, and recording quality generally, see Garbage In, Garbage Out.
What to ask a vendor
- Does your accuracy number cover speaker identification? The standard accuracy measure counts words. The transcript is compared against a correct one, and every word heard wrong, dropped or added that was never spoken counts as an error, so up to 95% accuracy means about five such errors in every hundred words. A line with every word right and the wrong name on it counts as no errors at all. Speaker identification has a separate measure, so ask whether the vendor uses it. A vendor who has never separated the two has not measured who said what.
- Will you run my hardest multi-speaker recording, not your cleanest sample? A group interview over a phone line, a hearing with interruptions. The recordings that worry you are the real test.
- What happens when a name cannot be confirmed? A generic label such as Speaker 2 is more honest than a guess, and a vendor should say which they do.
- On which service do the speaker labels get corrected? Transcription is sold as two things: an automated draft, which labels speakers generically, and a verified record, in which the speakers have been told apart and named from what the recording supports. Ask which you are being quoted for.
Where VIQ comes in
VIQ’s automated drafts label each speaker, with generic labels where no names are supplied. Its verified records have the speakers told apart and named from the recording and its context. If the recordings that matter to you are the ones with several people in the room, that is the difference to ask about.
