Set up output scores

Every score the API returns is a number between 0 and 100, where higher means the model is more certain. There are two families:

Frame-level probability scores — well-calibrated probabilities sampled on a regular time axis, describing what is happening in the audio at each instant.
Confidence scores — how certain the model is about a speaker assignment, per turn.

Which scores are available

Score	Request parameter	Shape
Speech probability	`speechProbability`	One curve
Cross-talk probability	`crosstalkProbability`	One curve
Speaker probability	`speakerProbability`	One curve per speaker
Turn-level confidence	`turnLevelConfidence`	Per diarization segment
Identification confidence	(always on)	Per voiceprint match

All of these work the same way on diarize and identify.

Frame-level probability scores

The three probability scores answer three different questions about the same moment in time:

speechProbability — is anyone speaking?
crosstalkProbability — is more than one person speaking?
speakerProbability — is this particular speaker speaking?

Each is a boolean request parameter, defaulting to false. Each returns an object with a score and a resolution:

Example output

{
  "speechProbability": {
    "resolution": 0.02,
    "score": [95, 89, 78, 67, 56, 45, 34, 23, 12, 1]
  },
  "crosstalkProbability": {
    "resolution": 0.02,
    "score": [3, 8, 61, 74, 12, 4, 2, 1, 0, 0]
  },
  "speakerProbability": {
    "resolution": 0.02,
    "score": {
      "SPEAKER_00": [95, 89, 12, 4, 8, 2, 1, 0, 0, 0],
      "SPEAKER_01": [5, 11, 88, 91, 85, 44, 33, 23, 12, 1]
    }
  }
}

score — for speechProbability and crosstalkProbability, an array of values between 0 and 100. For speakerProbability, an object keyed by speaker label, each holding such an array.
resolution — seconds per sample. Fixed at 0.02 (20ms).

All three share the same frame axis, so index i refers to the same 20ms window in every score. To convert an index to a timestamp, multiply by the resolution: sample i starts at i * resolution seconds.

On identification jobs, speakerProbability is always keyed by diarization labels (SPEAKER_00, SPEAKER_01, …) — never by the voiceprint or target name. Use the identification array to map a diarization label to a matched name.

Example request

output_scores.py

import requests

data = {
    "url": "https://files.pyannote.ai/marklex1min.wav",
    "model": "precision-3",
    "speechProbability": True,
    "crosstalkProbability": True,
    "speakerProbability": True,
}

response = requests.post(
    "https://api.pyannote.ai/v1/diarize",
    headers={"Authorization": "Bearer YOUR_API_KEY", "Content-Type": "application/json"},
    json=data,
)
print(response.json())

curl -X POST "https://api.pyannote.ai/v1/diarize" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://files.pyannote.ai/marklex1min.wav",
    "model": "precision-3",
    "speechProbability": true,
    "crosstalkProbability": true,
    "speakerProbability": true
  }'

const data = {
  url: "https://files.pyannote.ai/marklex1min.wav",
  model: "precision-3",
  speechProbability: true,
  crosstalkProbability: true,
  speakerProbability: true,
};

const response = await fetch("https://api.pyannote.ai/v1/diarize", {
  method: "POST",
  headers: { Authorization: "Bearer YOUR_API_KEY", "Content-Type": "application/json" },
  body: JSON.stringify(data),
});
console.log(await response.json());

What to use the scores for

The scores exist so the decision can belong to your product logic rather than to the model.

Building high-quality voiceprints — keep only segments above a speakerProbability or turnLevelConfidence threshold, trading recall for precision.
Routing to human review — flag segments below a threshold instead of reviewing a whole transcript.
Monitoring quality over time — track probability distributions per audio source to catch drift before it reaches your users.
Resolving overlap for speech-to-text — most STT engines transcribe one speaker at a time. Use speakerProbability to decide which speaker is most likely being transcribed, or let exclusive diarization do it for you.
Extracting clean training data — raise crosstalkSensitivity to over-detect overlap (or filter on crosstalkProbability) to remove every contaminated section.

Confidence scores

Confidence scores measure how certain the model is about a speaker assignment, rather than about acoustic activity. There are two kinds.

Turn-level confidence

Add "turnLevelConfidence": true to your request. Each diarization segment gains a confidence object, keyed by speaker, giving the confidence of that speaker assignment:

{
  "start": 10.0,
  "end": 15.0,
  "speaker": "SPEAKER_00",
  "confidence": { "SPEAKER_00": 85, "SPEAKER_01": 15 }
}

Identification confidence

Identification confidence is specific to speaker identification and shows how well each voiceprint matches each speaker. It is included automatically when using the identify endpoint — no parameter required.

{
  "voiceprints": [
    {
      "speaker": "SPEAKER_00",
      "match": "John Doe",
      "confidence": {
        "John Doe": 86,
        "Jane Smith": 12
      }
    }
  ]
}

Identification confidence scores are different from diarization confidence scores. They measure how well a voiceprint matches a speaker, not how certain the diarization model is about the speaker assignment.

Getting Started

Tutorials

Webhooks

Support

Administration

Set up output scores

Which scores are available

Frame-level probability scores

Example request

What to use the scores for

Confidence scores

Turn-level confidence

Identification confidence

​Which scores are available

​Frame-level probability scores

​Example request

​What to use the scores for

​Confidence scores

​Turn-level confidence

​Identification confidence

Which scores are available

Frame-level probability scores

Example request

What to use the scores for

Confidence scores

Turn-level confidence

Identification confidence