Tune VAD and cross-talk sensitivity

Two input parameters let you trade precision against recall on the detection stages of the pipeline. Both are floats in [-5.0, 5.0] and both default to 0.0, which is the balanced setting.

Parameter	Controls	Higher values	Lower values
`vadSensitivity`	Voice activity detection	Increase recall — less speech is missed	Increase precision — less non-speech is detected as speech
`crosstalkSensitivity`	Overlapping speech detection	Detect more overlapping speech	Report only clear overlaps

The defaults are tuned for general-purpose audio. Reach for these parameters when you know something specific about your recordings that the balanced setting does not account for.

When to raise or lower `vadSensitivity`

Raise it when speech is being missed — quiet or distant speakers, soft back-channels (“mm-hm”, “right”), heavily compressed phone audio, or a recording where short utterances matter and you would rather over-detect than lose them. Lower it when non-speech is being picked up as speech — background music, keyboard noise, laughter, HVAC hum, or a noisy open-plan office. This is also the setting to reach for when downstream transcription is producing text for segments that contain no words.

When to raise or lower `crosstalkSensitivity`

Raise it for genuinely conversational audio where interruptions carry meaning: debates, multi-party meetings, contact-center calls where agent and customer talk over each other. Lower it when you need clean, non-overlapping segments — for example when feeding results to a transcription step that expects one speaker at a time, or when brief acknowledgements are being reported as overlap and fragmenting your turns.

Tune one parameter at a time and compare against a fixed sample of your own audio.

Two ways to handle overlap

There are two distinct approaches, depending on what you need. If you are interested in speaker diarization, use crosstalkSensitivity to bias detection toward or away from overlap, and read the resulting turns. If you are a power user interested in overlapping speech itself, request crosstalkProbability and apply your own threshold after the fact:

{
  "url": "https://files.pyannote.ai/marklex1min.wav",
  "model": "precision-3",
  "crosstalkProbability": true
}

Example request

sensitivity.py

import requests

data = {
    "url": "https://files.pyannote.ai/marklex1min.wav",
    "model": "precision-3",
    "vadSensitivity": 1.5,
    "crosstalkSensitivity": 1.5,
}

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",
    "vadSensitivity": 1.5,
    "crosstalkSensitivity": 1.5
  }'

const data = {
  url: "https://files.pyannote.ai/marklex1min.wav",
  model: "precision-3",
  vadSensitivity: 1.5,
  crosstalkSensitivity: 1.5,
};

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());

Both parameters are also accepted on the identify endpoint, with the same ranges and defaults.

Output scores — inspect the effect of these parameters frame by frame
Configuring the number of speakers — constrain how many speakers the model may find

Getting Started

Tutorials

Webhooks

Support

Administration

Tune VAD and cross-talk sensitivity

When to raise or lower `vadSensitivity`

When to raise or lower `crosstalkSensitivity`

Two ways to handle overlap

Example request

​When to raise or lower vadSensitivity

​When to raise or lower crosstalkSensitivity

​Two ways to handle overlap

​Example request

​Related

When to raise or lower `vadSensitivity`

When to raise or lower `crosstalkSensitivity`

Two ways to handle overlap

Example request

Related