Last updated: September 12, 2026
Why a term out of speech research lands on your scorecard
Speaker diarization splits a recording into turns and labels which turn belongs to the agent and which belongs to the caller. Transcription writes down the words. Diarization decides whose mouth they came from, which is the difference between a scorecard that grades your agent and one that grades the caller by mistake.
Take a criterion most floors carry somewhere on the card: acknowledged the caller's situation before moving to the fix. The system finds "that sounds really frustrating" at 00:21, matches it against the empathy set, and awards six points. If the caller said it, an agent got paid for the caller's manners.
Worse on a leasing call, where a caller volunteering that she has two young kids and an agent asking about them are opposite records under fair housing. One swapped label turns a clean call into evidence for the thing the program exists to prevent.
What diarization does, and how it differs from transcription
Diarization answers one question about an audio file: who spoke when. Speaker 1 holds the floor from 00:00 to 00:04, Speaker 2 takes over, and out comes a list of segments with a speaker label on each.
Notice what is missing. The system does not know that Speaker 1 is Dana on the overnight board, only that two voices are present and where each one talks. Attaching a real name is speaker identification, a separate job needing an enrolled voice print. In call QA, whichever side gives the greeting is the agent.
Transcription is the other half, and it fails independently. ASR turns waveform into words while diarization draws the boundaries on the same audio, so a transcript can be word perfect and still hand four sentences to the wrong person. Wrong words is a vocabulary problem, solved with phrase sets for greetings and an account's do-not-say terms. Boundaries landing in the wrong place is an audio problem, and no vocabulary work touches it.
Two channels or one, which decides most of this
Recordings arrive in two shapes, and the difference between them is bigger than anything else on this page.
A two channel recording, sometimes called dual channel or stereo, keeps the agent on one channel and the caller on the other. Your phone system did the separation while the call was live, because it knew which leg was which. Attribution stops being a prediction and becomes a fact stored in the file. Overtalk and two similar voices stop being problems, since the speaker is a channel number.
A mixed mono recording sums both voices into one waveform. The separation then has to be worked backwards out of the sound, clustering stretches of audio by voice characteristics and guessing where one speaker stops. That works. It works less well than reading a channel number, and that is true of every vendor in this category.
This is a real constraint in CallSniper today rather than a hypothetical one. Speaker separation holds up on two channel audio and gets unreliable on mixed mono, and the work to close that gap sits on our side. Any criterion written around who said a line inherits it.
The tidy version is to call mono audio unusable, and that overstates it badly. Most of a compliance card never touches speaker identity. Whether the recording disclosure arrived before the caller gave details, or the TPMO disclaimer landed before the benefits discussion, comes off a phrase and the timestamp under it, and that holds up on a mono file. The behavior half is what breaks. Who apologized, and who read the callback number back digit by digit.
Where diarization gets it wrong
Overtalk is the most common failure. Two voices at once produce audio belonging to both, the clustering has to pick one, and an interruption or a caller finishing the agent's sentence both land here.
Short turns come next. A one word "yeah" is barely enough audio to characterize a voice from, so it gets dropped or handed to whoever was talking around it. Two voices in the same pitch range cause a third kind of error, especially over a phone line that already threw away the high frequencies.
Audio quality carries the rest, since recordings get stored at 8 kHz and heavily compressed, with hold music and floor noise under the conversation. A third voice arriving on a warm transfer or a supervisor barge is worse again, because now the clustering has to settle how many speakers exist first.
What a confidence score means here
Every scored line in CallSniper carries a timestamp and a confidence value. Confidence is the system saying how sure it is about that line, which is different from how correct it turned out to be. High confidence on a rule based criterion can be taken at face value, and a low confidence flag is an instruction to go listen. Machine scoring holds up on rule based criteria and gets weaker on tone and empathy, which Aircall and StackAI both report about automated scoring generally.
Diarization has a measure of its own, called diarization error rate, or DER. It rolls three mistakes into one percentage of total audio time: speech the system missed, audio it marked as speech when nobody was talking, and time handed to the wrong speaker.
Be careful with a quoted DER figure. A rate measured on two people in a quiet room says little about 8 kHz audio of a caller on speakerphone in a truck, so ask what it looks like on audio shaped like yours.
What a swapped label costs on a scorecard
Say a card carries 18 criteria and 7 of them turn on who spoke. A speaker error does not take out the score. It lands on those 7 quietly, because the transcript still reads perfectly well with the labels swapped.
Bad attribution produces a confident score rather than an error. An agent can pick up six points for the caller's empathy line, drop four on a probing question the caller asked, and land near the right total by accident.
So while your audio is mono, treat speaker dependent criteria as a finding aid, where the flag points a supervisor at a timestamp. When disputes cluster on the same few rows, check attribution before rewriting the criterion, because the two look identical on a report. How a score is produced covers the rest of the chain, with the surrounding terms in the call center glossary.
Check your own recordings before you trust anybody's score
Find out what your phone system exports before you evaluate any vendor here, because the answer changes what scoring can do for you whoever you buy from.
Download one recording and open it in a free audio editor. A two channel file shows two stacked waveforms, and if those look different from each other, the two sides really are separated. Identical waveforms mean a mono recording saved inside a stereo container. Anyone technical can run ffprobe your-call.wav, which prints the channel count.
Then ask your provider, since the setting goes by a different name on every platform. Say dual channel recording, stereo recording, and separate legs. Plenty of systems ship with it switched off to save storage, and some cannot do it at all, which you want to know on a Tuesday rather than halfway through a rollout.
A scored sample of your own calls scores six months of your own recordings against the scorecard you already use, at a fixed pilot price, and every scored line opens as the transcript beside the audio. Bring a mono batch and a two channel batch if you have both, because the gap between them is the most useful thing that pilot can show you.
Frequently asked questions
What is speaker diarization?
Speaker diarization works out who spoke when in a recording, splitting audio into turns and tagging each turn as a distinct voice. Attaching a real name to one is a separate job.
What is the difference between diarization and speaker recognition?
Diarization separates voices without knowing whose they are, which is why its output says Speaker 1 and Speaker 2. Recognition matches a voice against an enrolled voice print to name the person.
What is diarization in ASR or speech to text?
It is the speaker layer on top of the words. ASR produces the text, diarization produces the boundaries, and a transcript with agent and caller turns is the two merged.
What is diarization error rate?
DER measures how wrong a diarization was, combining missed speech, audio wrongly marked as speech, and time given to the wrong speaker into one percentage of the audio. A benchmark figure will not carry over to phone calls.
Does speaker diarization work on mono recordings?
It runs, and it is less reliable than on two channel audio, because the separation gets inferred from the sound instead of read off a channel. Check what your phone system exports before building a scorecard around who said what.
Sources
- Transcript timestamps, per line confidence values, and speaker separation. CallSniper product documentation. Verified internally.
- Speaker attribution on single channel recordings. Two channel audio separates the two parties reliably and mixed mono does not. Internal, current status unverified.
- Diarization error rate as missed speech, false alarm speech, and speaker confusion. Speech research literature. Unverified.
- 8 kHz narrowband as a common call recording format. Unverified.
- Aircall and StackAI on rule based criteria scoring more reliably than tone. Third party, unverified.
- The 18 criterion card and the empathy line at 00:21. Invented for this article.